A deep-learning framework for brain tumor segmentation via three-dimensional mass-preserving geometric transformation Authors: Tsung-Ming Huang, Kai-Qian Zheng, Wen-Wei Lin, Tiexiang Li, Shing-Tung Yau Published: 2026 ABSTRACT This article presents a robust and efficient framework for brain tumor segmentation based on deep learning. We introduce a novel three-dimensional (3D) mass-preserving geometric transformation (MPGT) that employs a homotopy method to transform irregular brain magnetic resonance (MR) images into standardized solid cubes. This transformation preserves local mass ratios while maintaining global structural integrity, providing a structured input for deep learning models. Furthermore, we propose a modified two-phase segmentation strategy to minimize inference time and a postprocessing technique to enhance lesion-wise performance. Extensive validation on the Brain Tumor Segmentation (BraTS) Challenge 2023 dataset demonstrates that our method, when integrated with nnU-Net, achieves competitive Dice scores of 0.9282 (Whole Tumor), 0.8812 (Tumor Core), and 0.8527 (Enhanced Tumor). These results are superior to or comparable with top-ranking competition entries. Introduction Gliomas are the most common type of malignant brain tumor in adults. According to the World Health Organization (WHO), brain tumors can be classified into four grades, from 1 to 4, on the basis of their histological characteristics and molecular features. Grade 1 tumors are typically slow-growing and benign, whereas grade 4 tumors, including glioblastomas (GBMs), are the most aggressive and malignant forms. GBMs are among the deadliest types of cancer due to their location and invasive nature. The mutation status of isocitrate dehydrogenase (IDH) is considered the most critical marker for glioma diagnosis [1]. Deep learning techniques [2–4] have been shown to accurately and objectively predict the status of IDH mutations on the basis of magnetic resonance imaging (MRI) data. However, in most studies involving IDH mutations and deep learning, glioma regions must be segmented along the lesion edges [5]. Thus, detecting and segmenting brain tumors on the basis of MRI data are essential tasks. The Brain Tumor Segmentation (BraTS) Challenge [6–10] serves as a robust platform that offers a publicly available dataset for model development. The objective of this competition is to evaluate the latest and most advanced methods for segmenting diffuse gliomas in the brain. The BraTS Challenge 2023 includes 2040 baseline 3D MR images of adult gliomas, with the aim of segmenting the whole tumor (WT), tumor core (TC), and enhanced tumor (ET) regions. Each MR image features four modalities: native T1, post-contrast T1-weighted (T1Gd), T2-weighted (T2) and fluid-attenuated inversion recovery (T2-FLAIR). Each modality is represented as a tensor image with dimensions of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$240 \times 240 \times 155$$\end{document}. In the past decade, U-Net-based deep learning architectures, such as ResUNet [11], SegResNet [12], nnU-Net [13], and MedNeXt [14], have emerged as crucial and effective methods for brain tumor segmentation. Luu et al. [15] utilized nnU-Net as a baseline, integrating BraTS-specific optimizations from [16] along with their own enhancements to accurately segment tumor regions. Similarly, [17] introduced an ensemble approach combining DeepSeg [18], the improved nnU-Net [15], and DeepSCAN [19] for the automatic detection of glioma boundaries in preoperative MR images. Building upon these foundational approaches, the field has seen rapid progress, with more sophisticated models and strategies emerging, particularly evident in works from 2024–2025. Noteworthy recent contributions include enhanced multi-modal multi-scale contextual aggregation with attention fusion for improved segmentation [20], the application of YOLO-based deep learning for automated brain tumor segmentation and classification [21], continued advancements in U-Net based methods through transfer learning for segmentation enhancement [22], and comprehensive deep learning frameworks that address both 3D segmentation and survival prediction [20]. For instance, Ferreira et al. [23], a top-performing entry in the BraTS 2023 challenge, utilized generative adversarial networks and registration techniques to significantly expand training data for nnU-Net and Swin UNETR [24] algorithms. Their approach involved proposing three methods for synthetic data generation and combining all inferred results to predict tumor regions. Maani et al. [25], another strong participant in BraTS 2023, leveraged SegResNet and MedNeXt with deep supervision and sophisticated post-processing to achieve competitive results. Kasliwal et al. [26] recently introduced ReFuSeg, a regularized multi-modal fusion model specifically designed to achieve precise brain tumor segmentation even with missing modalities and in the presence of imaging artifacts. These works collectively highlight the continuous drive for more robust and generalizable segmentation methods. However, limitations in GPU memory remain challenging owing to the large size of each BraTS image, with dimensions of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$240 \times 240 \times 155$$\end{document}. To address this issue, random cropping of the raw brain images is commonly employed; however, this approach may not capture all relevant information from the original data. To better preserve the global features of the MRI data, a two-stage optimal mass transport (OMT) method with the mass-preserving parameterization was proposed in [27]. In the first stage, the irregular brain domain in the MR image is transformed into a solid unit sphere using the mass-preserving OMT \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f_b$$\end{document}. In the second stage, the cubic domain is mapped to a unit sphere using the volume-preserving OMT \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f_c$$\end{document}. The combined OMT map \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f = f_c^{-1} \circ f_b$$\end{document} can subsequently be used to convert the irregular brain domain into a cubic form. Building on this OMT approach, Lin et al. [28] introduced a novel two-phase residual U-Net algorithm for detecting tumor regions. In Phase I, the authors employed the residual U-Net with OMT tensors to accurately predict the WT region. In Phase II, they generated a smooth density function based on the predicted WT region, thereby enlarging the potential tumor region in the new OMT tensors, which serve as inputs for the residual U-Net to effectively train the model. The geometric transformation achieved through the mass-preserving OMT approach serves as a critical framework for data preprocessing in deep learning applications for brain tumor segmentation. The aim of this paper is to enhance (i) the efficiency of the mass-preserving geometric transformation (MPGT) and the two-phase segmentation method and (ii) the lesion-wise Dice scores through a proposed postprocessing step. Ultimately, the goal is to establish a more robust and efficient deep learning framework for brain tumor segmentation. The three main contributions of this paper can be summarized as follows. Efficient Geometric Transformation: We propose a homotopy-based mass-preserving map that limits inversion calculations to spherical surfaces, significantly reducing computational cost compared to traditional Optimal Mass Transport (OMT) methods [29, 30]. We provide a comparative analysis of runtime to substantiate these efficiency gains.Two-phase Segmentation Strategy: We introduce a pipeline that utilizes randomly cropped raw data for initial tumor localization (Phase I) and MPGT tensors for refined segmentation (Phase II), balancing inference speed with global context preservation.Enhanced Post-processing: We implement a threshold-based filtering mechanism derived from training data distributions to reduce false positives and improve lesion-wise Dice scores.This paper is organized as follows. Section 2 introduces the mass-preserving parameterization and our proposed homotopy method. In Sect. 3, we present a two-phase segmentation method that incorporates mass-preserving geometric transformation and postprocessing steps. Section 4 provides numerical results demonstrating the effectiveness of our proposed two-phase segmentation approach. Finally, we present our conclusions in Sect. 5. Efficient Geometric Transformation: We propose a homotopy-based mass-preserving map that limits inversion calculations to spherical surfaces, significantly reducing computational cost compared to traditional Optimal Mass Transport (OMT) methods [29, 30]. We provide a comparative analysis of runtime to substantiate these efficiency gains. Efficient Geometric Transformation: We propose a homotopy-based mass-preserving map that limits inversion calculations to spherical surfaces, significantly reducing computational cost compared to traditional Optimal Mass Transport (OMT) methods [29, 30]. We provide a comparative analysis of runtime to substantiate these efficiency gains. Two-phase Segmentation Strategy: We introduce a pipeline that utilizes randomly cropped raw data for initial tumor localization (Phase I) and MPGT tensors for refined segmentation (Phase II), balancing inference speed with global context preservation. Two-phase Segmentation Strategy: We introduce a pipeline that utilizes randomly cropped raw data for initial tumor localization (Phase I) and MPGT tensors for refined segmentation (Phase II), balancing inference speed with global context preservation. Enhanced Post-processing: We implement a threshold-based filtering mechanism derived from training data distributions to reduce false positives and improve lesion-wise Dice scores. Enhanced Post-processing: We implement a threshold-based filtering mechanism derived from training data distributions to reduce false positives and improve lesion-wise Dice scores. 3D mass-preserving geometric transformation OMT with volume-/mass-preserving maps [29, 30] has proven to be an effective approach for detecting and segmenting brain tumors [27, 28, 31]. Using OMT, the irregular 3D brain MR image can be transformed into a 3D solid cube \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {C}$$\end{document}, depicted in Fig. 1, which helps to preserve the global information in the MR image. In this section, we propose a modified mass-preserving geometric transformation via the homotopy method to reduce the computational cost of generating the cube \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {C}$$\end{document}. We use the MPGT (Mass-Preserving Geometric Transformation) tensors, generated from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {C}$$\end{document}, as inputs to train our learning model and subsequently infer MPGT tensors for the validation and testing data. By converting the inference results obtained on the basis of the MPGT tensors back to the original brain images, we get the final segmentation of the tumor region.Fig. 1Using the OMT method to transform the manifold \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {M}$$\end{document} into the cube \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {C}$$\end{document} Using the OMT method to transform the manifold \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {M}$$\end{document} into the cube \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {C}$$\end{document} Overview of the transformation pipeline Our goal is to convert each irregular brain volume into a standardized cubic tensor that (i) preserves the global anatomy and (ii) allocates more voxels to tumor regions, so that the downstream network can better detect small lesions. Conceptually, the proposed 3D mass-preserving geometric transformation (MPGT) consists of the following stages: Start from the original brain mesh. We represent the brain volume as a tetrahedral mesh \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {M}$$\end{document} with a scalar density \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rho _\gamma $$\end{document} defined on each tetrahedron. Intuitively, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rho _\gamma $$\end{document} encodes how much “importance” (mass) we assign to each local region; tumor voxels are assigned higher density so that they will be expanded later in the cubic domain.Normalize the boundary via spherical area-preserving maps. The brain boundary \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\partial \mathcal {M}$$\end{document} has an irregular, patient-specific shape. We first map \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\partial \mathcal {M}$$\end{document} onto the unit sphere \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\partial \mathcal {B}^3$$\end{document} with an area-measure-preserving map. In parallel, we also map the boundary of a reference cube \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\partial \mathcal {C}$$\end{document} onto \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\partial \mathcal {B}^3$$\end{document}. These two spherical maps (i) remove anatomical variations in head shape by placing all brains into a common spherical boundary, and (ii) allow us to construct a direct, area-preserving correspondence between \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\partial \mathcal {M}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\partial \mathcal {C}$$\end{document} by composing the two boundary maps.Fill the interior with a homotopy-driven volumetric map. Once the boundaries of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {M}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {C}$$\end{document} are matched on the sphere, we deform the interior of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {M}$$\end{document} into the cube \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {C}$$\end{document} by gradually moving boundary vertices from the identity map to the target boundary map and solving a sequence of linear systems (the homotopy method). At each step, we update an OMT-type Laplacian and recompute interior vertices so that local volumes are transported in a mass-preserving way and the mapping remains close to bijective with negligible folding.Effect on the tumor in the cubic tensor. Because tumor regions have larger \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rho _\gamma $$\end{document}, the map allocates a disproportionately larger volume to them in the cube: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned}&|f_\gamma (\tau )| > |\tau | \quad \text {for tumor tetrahedra } \tau , \qquad \\&|f_\gamma (\tau )| \approx |\tau | \quad \text {for non-tumor tetrahedra.} \end{aligned}$$\end{document} After rasterization, this implies that tumors occupy more voxels in the final \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$m \times m \times m$$\end{document} tensor, which increases their visibility to the network, especially for small foci, while preserving the global brain context.Why this helps segmentation and efficiency. For segmentation quality, the network sees a globally aligned, cube-shaped input in which tumors are geometrically emphasized while the overall anatomy is preserved. For computational efficiency, instead of solving a fully nonlinear volumetric OMT problem over the entire 3D domain, we restrict the more expensive computations to the spherical boundary and then propagate them inward via homotopy, thereby reducing runtime compared to traditional volumetric OMT.In the following subsections, we formalize this pipeline. Section 2.2 introduces the mass-preserving parameterization and its associated energy, while Sect. 2.3 details the homotopy construction that efficiently connects the area-preserving boundary maps to the volumetric mass-preserving map used to generate the MPGT tensors. Start from the original brain mesh. We represent the brain volume as a tetrahedral mesh \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {M}$$\end{document} with a scalar density \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rho _\gamma $$\end{document} defined on each tetrahedron. Intuitively, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rho _\gamma $$\end{document} encodes how much “importance” (mass) we assign to each local region; tumor voxels are assigned higher density so that they will be expanded later in the cubic domain. Start from the original brain mesh. We represent the brain volume as a tetrahedral mesh \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {M}$$\end{document} with a scalar density \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rho _\gamma $$\end{document} defined on each tetrahedron. Intuitively, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rho _\gamma $$\end{document} encodes how much “importance” (mass) we assign to each local region; tumor voxels are assigned higher density so that they will be expanded later in the cubic domain. Normalize the boundary via spherical area-preserving maps. The brain boundary \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\partial \mathcal {M}$$\end{document} has an irregular, patient-specific shape. We first map \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\partial \mathcal {M}$$\end{document} onto the unit sphere \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\partial \mathcal {B}^3$$\end{document} with an area-measure-preserving map. In parallel, we also map the boundary of a reference cube \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\partial \mathcal {C}$$\end{document} onto \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\partial \mathcal {B}^3$$\end{document}. These two spherical maps (i) remove anatomical variations in head shape by placing all brains into a common spherical boundary, and (ii) allow us to construct a direct, area-preserving correspondence between \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\partial \mathcal {M}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\partial \mathcal {C}$$\end{document} by composing the two boundary maps. Normalize the boundary via spherical area-preserving maps. The brain boundary \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\partial \mathcal {M}$$\end{document} has an irregular, patient-specific shape. We first map \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\partial \mathcal {M}$$\end{document} onto the unit sphere \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\partial \mathcal {B}^3$$\end{document} with an area-measure-preserving map. In parallel, we also map the boundary of a reference cube \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\partial \mathcal {C}$$\end{document} onto \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\partial \mathcal {B}^3$$\end{document}. These two spherical maps (i) remove anatomical variations in head shape by placing all brains into a common spherical boundary, and (ii) allow us to construct a direct, area-preserving correspondence between \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\partial \mathcal {M}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\partial \mathcal {C}$$\end{document} by composing the two boundary maps. Fill the interior with a homotopy-driven volumetric map. Once the boundaries of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {M}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {C}$$\end{document} are matched on the sphere, we deform the interior of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {M}$$\end{document} into the cube \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {C}$$\end{document} by gradually moving boundary vertices from the identity map to the target boundary map and solving a sequence of linear systems (the homotopy method). At each step, we update an OMT-type Laplacian and recompute interior vertices so that local volumes are transported in a mass-preserving way and the mapping remains close to bijective with negligible folding. Fill the interior with a homotopy-driven volumetric map. Once the boundaries of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {M}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {C}$$\end{document} are matched on the sphere, we deform the interior of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {M}$$\end{document} into the cube \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {C}$$\end{document} by gradually moving boundary vertices from the identity map to the target boundary map and solving a sequence of linear systems (the homotopy method). At each step, we update an OMT-type Laplacian and recompute interior vertices so that local volumes are transported in a mass-preserving way and the mapping remains close to bijective with negligible folding. Effect on the tumor in the cubic tensor. Because tumor regions have larger \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rho _\gamma $$\end{document}, the map allocates a disproportionately larger volume to them in the cube: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned}&|f_\gamma (\tau )| > |\tau | \quad \text {for tumor tetrahedra } \tau , \qquad \\&|f_\gamma (\tau )| \approx |\tau | \quad \text {for non-tumor tetrahedra.} \end{aligned}$$\end{document} After rasterization, this implies that tumors occupy more voxels in the final \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$m \times m \times m$$\end{document} tensor, which increases their visibility to the network, especially for small foci, while preserving the global brain context. Effect on the tumor in the cubic tensor. Because tumor regions have larger \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rho _\gamma $$\end{document}, the map allocates a disproportionately larger volume to them in the cube: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned}&|f_\gamma (\tau )| > |\tau | \quad \text {for tumor tetrahedra } \tau , \qquad \\&|f_\gamma (\tau )| \approx |\tau | \quad \text {for non-tumor tetrahedra.} \end{aligned}$$\end{document} After rasterization, this implies that tumors occupy more voxels in the final \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$m \times m \times m$$\end{document} tensor, which increases their visibility to the network, especially for small foci, while preserving the global brain context. Why this helps segmentation and efficiency. For segmentation quality, the network sees a globally aligned, cube-shaped input in which tumors are geometrically emphasized while the overall anatomy is preserved. For computational efficiency, instead of solving a fully nonlinear volumetric OMT problem over the entire 3D domain, we restrict the more expensive computations to the spherical boundary and then propagate them inward via homotopy, thereby reducing runtime compared to traditional volumetric OMT. Why this helps segmentation and efficiency. For segmentation quality, the network sees a globally aligned, cube-shaped input in which tumors are geometrically emphasized while the overall anatomy is preserved. For computational efficiency, instead of solving a fully nonlinear volumetric OMT problem over the entire 3D domain, we restrict the more expensive computations to the spherical boundary and then propagate them inward via homotopy, thereby reducing runtime compared to traditional volumetric OMT. Mass-preserving parameterization Let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {M}$$\end{document} be a discrete simplicial manifold representing an irregular 3D brain image with a genus-zero boundary. This tetrahedral mesh consists of sets of vertices \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbb {V}(\mathcal {M})$$\end{document}, edges \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbb {E}(\mathcal {M})$$\end{document}, faces \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbb {F}(\mathcal {M})$$\end{document}, and tetrahedrons \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbb {T}(\mathcal {M})$$\end{document}. Let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$| \alpha |$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$| \textbf{v} |$$\end{document} denote the area of the triangular mesh \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha $$\end{document} and the volume of the tetrahedron \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{v}$$\end{document}, respectively. Step-by-step summary of the parameterization. Before presenting the formal definitions, we briefly summarize the construction: We start from a tetrahedral mesh \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {M}$$\end{document} equipped with a density function \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rho _\gamma $$\end{document} on each tetrahedron.We seek a piecewise linear map \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f: \mathcal {M} \rightarrow \mathcal {B}^3$$\end{document} such that the volume of each image tetrahedron \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f(\tau )$$\end{document} is proportional to the prescribed mass \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rho _\gamma (\tau )|\tau |$$\end{document}.This requirement is encoded via the volume-weighted stretch energy \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$E_V(f)$$\end{document}; minimizing \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$E_V$$\end{document} under a global volume constraint enforces local mass preservation.We fix the boundary of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {M}$$\end{document} on the unit sphere and solve the resulting constrained optimization problem. Theorem 1 then shows that the minimizer of this problem is exactly the desired mass-preserving map.The detailed mathematical formulation of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F_{\rho _\gamma }$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$E_V(f)$$\end{document} is given below. We start from a tetrahedral mesh \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {M}$$\end{document} equipped with a density function \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rho _\gamma $$\end{document} on each tetrahedron. We start from a tetrahedral mesh \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {M}$$\end{document} equipped with a density function \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rho _\gamma $$\end{document} on each tetrahedron. We seek a piecewise linear map \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f: \mathcal {M} \rightarrow \mathcal {B}^3$$\end{document} such that the volume of each image tetrahedron \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f(\tau )$$\end{document} is proportional to the prescribed mass \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rho _\gamma (\tau )|\tau |$$\end{document}. We seek a piecewise linear map \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f: \mathcal {M} \rightarrow \mathcal {B}^3$$\end{document} such that the volume of each image tetrahedron \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f(\tau )$$\end{document} is proportional to the prescribed mass \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rho _\gamma (\tau )|\tau |$$\end{document}. This requirement is encoded via the volume-weighted stretch energy \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$E_V(f)$$\end{document}; minimizing \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$E_V$$\end{document} under a global volume constraint enforces local mass preservation. This requirement is encoded via the volume-weighted stretch energy \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$E_V(f)$$\end{document}; minimizing \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$E_V$$\end{document} under a global volume constraint enforces local mass preservation. We fix the boundary of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {M}$$\end{document} on the unit sphere and solve the resulting constrained optimization problem. Theorem 1 then shows that the minimizer of this problem is exactly the desired mass-preserving map. We fix the boundary of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {M}$$\end{document} on the unit sphere and solve the resulting constrained optimization problem. Theorem 1 then shows that the minimizer of this problem is exactly the desired mass-preserving map. In [29, 30], a mass-preserving parameterization \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f \in \textbf{F}_{\rho _{\gamma }}$$\end{document} was proposed to transform a domain \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {M}$$\end{document}, equipped with a density function \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rho _{\gamma }$$\end{document}, into a unit sphere \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {B}^3 = \{ \textbf{x} \in \mathbb {R}^3; \Vert \textbf{x} \Vert _2 \le 1 \}$$\end{document} that possesses a uniform density \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rho _{\gamma }^b \equiv 1$$\end{document}. Here, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{F}_{\rho _{\gamma }}$$\end{document} denotes the set of all mass-preserving piecewise linear maps, defined as1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{F}_{\rho _{\gamma }} =&\, \{ f : \mathcal {M} \rightarrow \mathcal {B}^3 \,\big |\, \rho _{\gamma }(\tau ) | \tau | = \\&\quad\rho _{\gamma }^b(f(\tau )) | f(\tau ) | = | f(\tau ) |, \ \forall \tau \in \mathbb {T}(\mathcal {M}) \} . \end{aligned}$$\end{document}To compute the mass-preserving parameterization f, we define the volume-weighted stretch energy. This energy builds upon the work presented in [32], with its convergence properties further analyzed in [29], and is expressed as follows:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} {E}_{V}({f}) = \frac{1}{2}\text {trace}\left( \textbf{f}^\top L_{V}({f}) \, \textbf{f} \right) , \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {trace}(\cdot )$$\end{document} denotes the trace operator (the sum of diagonal elements), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{f}$$\end{document} is the matrix of vertex positions, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$L_{V}({f})$$\end{document} is the mass-weighted Laplacian matrix, given by 2a\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} {[L_{V}({f})]}_{i,j} = {\left\{ \begin{array}{ll} -w_{i,j}({f}) & \text {if } [\textbf{v}_i,\textbf{v}_j]\in \mathbb {E}(\mathcal {M}),\\ \sum _{\ell \ne i} w_{i,\ell }({f}) & \text {if } j = i,\\ 0 & \text {otherwise,} \end{array}\right. } \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$w_{i,j}({f})$$\end{document} is the stretched cotangent weight, defined as2b\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} &w_{i,j}(f) \\&\quad= -\frac{1}{9} \sum _{k,\ell } \frac{|f([\textbf{v}_i, \textbf{v}_k, \textbf{v}_\ell ])| |f([\textbf{v}_j, \textbf{v}_\ell , \textbf{v}_k])| \cos \theta _{i,j}^{k,\ell }(f)}{\rho (\tau )|\tau |}. \end{aligned}$$\end{document} Here, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _{i,j}^{k,\ell }(f)$$\end{document} is the dihedral angle between the triangles formed by \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$[\textbf{f}_i, \textbf{f}_k, \textbf{f}_{\ell }]$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$[\textbf{f}_j, \textbf{f}_{\ell }, \textbf{f}_k]$$\end{document}, with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{f}_t =f(\textbf{v}_t)=({f}_t^1, {f}_t^2, {f}_t^3)\in \mathbb {R}^3$$\end{document} for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t=1, \dots , n\equiv \#\mathbb {V}(\mathcal {M})$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{f}=[\textbf{f}_1^{\top }, \dots , \textbf{f}_n^{\top }]^{\top }\in \mathbb {R}^{n\times 3}$$\end{document}. An illustration for the dihedral angle \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _{i,j}^{k,\ell }(f)$$\end{document} is shown in Fig. 2. The stretched cotangent weight \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$w_{i,j}(f)$$\end{document} originates from the discretization of the Laplace-Beltrami operator on a triangulated surface and reflects the local geometric relationship between neighboring vertices. The cotangent terms encode the angles opposite to an edge in adjacent triangles, thereby capturing local curvature and mesh geometry, while the stretching factor accounts for the deformation induced by the transport map. Consequently, the weight measures how geometric information propagates across the mesh under the OMT-based metric.Fig. 2An illustration for the dihedral angle between \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$[\textbf{f}_i, \textbf{f}_k, \textbf{f}_\ell ]$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$[\textbf{f}_j, \textbf{f}_\ell , \textbf{f}_k]$$\end{document} in the tetrahedron \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$[\textbf{f}_i, \textbf{f}_j, \textbf{f}_{k}, \textbf{f}_{\ell }]$$\end{document} An illustration for the dihedral angle between \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$[\textbf{f}_i, \textbf{f}_k, \textbf{f}_\ell ]$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$[\textbf{f}_j, \textbf{f}_\ell , \textbf{f}_k]$$\end{document} in the tetrahedron \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$[\textbf{f}_i, \textbf{f}_j, \textbf{f}_{k}, \textbf{f}_{\ell }]$$\end{document} We say that the piecewise affine map \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f:\mathcal {M}\rightarrow \mathcal {B}^3$$\end{document} is induced by \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{f}$$\end{document}. The following theorem establishes the relationship between the mass-preserving parameterization and the volume-weighted stretch energy. Theorem 1 ([29] Let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {M}\subset \mathbb {R}^3$$\end{document} be a simply connected 3-manifold with a genus-zero boundary and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mu (\mathcal {M}) \equiv \sum _{\tau \in \mathbb {T}(\mathcal {M})}\rho _{\gamma }(\tau )|\tau |$$\end{document}. Let3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \mathcal {S}_f =&\, \{ E_V(f) \,| \, f:\mathcal {M}\rightarrow \mathbb {R}^3 \text { is}\,\\&\text{ orientation-preserving with} \sum _{\tau \in \mathbb {T}(\mathcal {M})}|f(\tau )|=\mu (\mathcal {M}) \} . \end{aligned}$$\end{document}Then, we have that (i)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\min \mathcal {S}_f\ge \frac{3}{2}\mu (\mathcal {M})$$\end{document};(ii)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f^* = \operatorname {argmin}\mathcal {S}_{f}$$\end{document} with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$E_V(f^*)= \frac{3}{2}\mu (\mathcal {M})$$\end{document} if and only if \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f^*$$\end{document} is mass preserving, i.e., \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$|f^*(\tau )| = \rho _{\gamma }(\tau ) | \tau |$$\end{document} for every \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tau \in \mathbb {T}(\mathcal {M})$$\end{document}. (i)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\min \mathcal {S}_f\ge \frac{3}{2}\mu (\mathcal {M})$$\end{document}; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\min \mathcal {S}_f\ge \frac{3}{2}\mu (\mathcal {M})$$\end{document}; (ii)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f^* = \operatorname {argmin}\mathcal {S}_{f}$$\end{document} with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$E_V(f^*)= \frac{3}{2}\mu (\mathcal {M})$$\end{document} if and only if \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f^*$$\end{document} is mass preserving, i.e., \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$|f^*(\tau )| = \rho _{\gamma }(\tau ) | \tau |$$\end{document} for every \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tau \in \mathbb {T}(\mathcal {M})$$\end{document}. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f^* = \operatorname {argmin}\mathcal {S}_{f}$$\end{document} with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$E_V(f^*)= \frac{3}{2}\mu (\mathcal {M})$$\end{document} if and only if \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f^*$$\end{document} is mass preserving, i.e., \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$|f^*(\tau )| = \rho _{\gamma }(\tau ) | \tau |$$\end{document} for every \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tau \in \mathbb {T}(\mathcal {M})$$\end{document}. A volumetric stretch energy minimization (VSEM) algorithm was proposed in [32] and further developed with projected gradient and homotopy techniques in [30] to find the minimizer of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S_f$$\end{document} in (3) with fixed spherical boundary points so that \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$| f(\mathcal {M}) | = \frac{4}{3} \pi $$\end{document} as follows. Let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{f}^* = [ (\textbf{f}^*_\texttt{I})^{\top }, (\textbf{f}^*_\texttt{B})^{\top } ]^{\top }$$\end{document} with\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \texttt{B} = \{t \mid \textbf{v}_t\in \partial \mathcal {M}\}\quad \text {and}\quad \texttt{I} = \{1, \ldots , n\} \backslash \texttt{B} \ne \varnothing . \end{aligned}$$\end{document}The Karush-Kuhn-Tucker (KKT) conditions of (3) under a given spherical boundary constraint \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{f}^{*}_{\texttt{B}}$$\end{document} lead to4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} {[}L_V(f^*)]_{\texttt{I},\texttt{I}} \textbf{f}^*_\texttt{I}= -[L_V(f^*)]_{\texttt{I},\texttt{B}} \textbf{f}^{*}_\texttt{B}. \end{aligned}$$\end{document}A fixed-point iterative method is then utilized to solve for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{f}^*_\texttt{I}$$\end{document} in (4). Homotopy method Given a density function \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rho _{\gamma }$$\end{document} for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {M}$$\end{document}, a mass-preserving OMT map \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f_{\rho _{\gamma }}^*$$\end{document} was proposed in [29, 30] to transform \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {M}$$\end{document} into the unit ball \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {B}^3 = \{ \textbf{x} \in \mathbb {R}^3; \Vert \textbf{x} \Vert _2 \le 1 \}$$\end{document} with a constant density function \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rho _{\gamma }^b \equiv 1$$\end{document}. The map \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f_{\rho _{\gamma }}^*$$\end{document} satisfies the following discrete OMT problem5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} f_{\rho _{\gamma }}^{*}=&\, \operatorname {argmin}_{f \in \textbf{F}_{\rho _{\gamma }}}\sum _{\textbf{v} \in \mathbb {V}(\mathcal {M})} \Vert \textbf{v} - f(\textbf{v}) \Vert _2^2 m_{\rho _{\gamma }}(\textbf{v}) \\\equiv&\, \operatorname {argmin}_{f \in \textbf{F}_{\rho _{\gamma }}} C_{\text{ v }}(f), \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{F}_{\rho _{\gamma }}$$\end{document} is defined in (1). The local volume measures \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$m_{\rho _{\gamma }}(\textbf{v})$$\end{document} at \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{v} \in \mathbb {V}(\mathcal {M})$$\end{document} are defined as\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} m_{\rho _{\gamma }}(\textbf{v}) = \frac{1}{4} \sum _{\tau \in N(\textbf{v})} \rho _{\gamma }(\tau ) | \tau |, \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$N(\textbf{v})$$\end{document} is the set of 1-ring neighboring tetrahedrons of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{v}$$\end{document}.Fig. 3A diagram illustrating the construction of the homotopy between \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(\mathcal {M}, \rho _{\gamma })$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(\mathcal {B}^3, \rho _{\gamma }^b \equiv 1)$$\end{document} A diagram illustrating the construction of the homotopy between \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(\mathcal {M}, \rho _{\gamma })$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(\mathcal {B}^3, \rho _{\gamma }^b \equiv 1)$$\end{document} A practical numerical method combined an area-measure-preserving OMT (AOMT) map with a homotopy method was proposed in [30] to compute OMT map \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f_{\rho _{\gamma }}^*$$\end{document}. Let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{f}^* = [ (\textbf{f}^*_\texttt{I})^{\top }, (\textbf{f}^*_\texttt{B})^{\top } ]^{\top }$$\end{document} be the vertex position matrix inducing of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f_{\rho _{\gamma }}^*$$\end{document}. As illustrated in Fig. 3, the process begins with an AOMT map \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ g_{\rho _{\gamma }}^* $$\end{document}, which transforms the boundary \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\partial \mathcal {M}$$\end{document} into \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\partial \mathcal {B}^3$$\end{document}6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} g_{\rho _{\gamma }}^{*}=&\, \operatorname {argmin}_{g \in \textbf{G}_{\rho _{\gamma }}} \sum _{\hat{\textbf{v}} \in \mathbb {V}(\partial \mathcal {M})} \Vert \hat{\textbf{v}} - g(\hat{\textbf{v}}) \Vert _2^2 a_{\rho _{\gamma }}(\hat{\textbf{v}}) \\ \equiv&\, \operatorname {argmin}_{g \in \textbf{G}_{\rho _{\gamma }}} C_{\text{ a }}(g), \end{aligned}$$\end{document}where\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{G}_{\rho _{\gamma }} = \left\{ g : \partial \mathcal {M} \rightarrow \partial \mathcal {B}^3 \,\right| \, \left. \rho _{\gamma }(\alpha ) | \alpha | = | g(\alpha ) |, \ \forall \alpha \in \mathbb {F}(\partial \mathcal {M}) \right\} \end{aligned}$$\end{document}and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$a_{\rho _{\gamma }}(\hat{\textbf{v}})$$\end{document} is the local area measure at \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{\textbf{v}} \in \mathbb {V}(\partial \mathcal {M})$$\end{document}, defined as\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} a_{\rho _{\gamma }}(\hat{\textbf{v}}) = \frac{1}{3} \sum _{\alpha \in N(\hat{\textbf{v}})} \rho _{\gamma }(\alpha ) |\alpha |. \end{aligned}$$\end{document}A projected gradient method is used to solve (6). Its convergence rate can be improved by applying FISTA [33], a variation of Nesterov’s accelerated gradient method [34]. Specifically, the basic gradient update\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned}& {\textrm{vec}}({\textbf{g}}^{(k+1)}) = {\textrm{vec}}(\textbf{g}^{(k)}) - \alpha _k \nabla {C}_a({\textrm{vec}}(\textbf{g}^{(k)})), \\& \alpha _k = \operatorname {argmin}_{\alpha >0} {C}_{\text{ a }}({\textrm{vec}}({\textbf{g}}^{(k+1)})) \end{aligned}$$\end{document}is replaced with the following updates in our new AOMT algorithm 7a\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} {\textrm{vec}}(\widehat{\textbf{g}}^{(k+1)})&= {\textrm{vec}}(\textbf{g}^{(k)}) - \alpha _k \nabla {C}_{\text{ a }}(\textbf{g}^{(k)}), \end{aligned}$$\end{document}7b\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} {\textrm{vec}}({\textbf{g}}^{(k+1)})&= (1-\beta _k) {\textrm{vec}}(\widehat{\textbf{g}}^{(k+1)}) + \beta _m {\textrm{vec}}(\widehat{\textbf{g}}^{(k)}), \end{aligned}$$\end{document}7c\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \beta _k&= \frac{1-\lambda _k}{\lambda _{k+1}},~ \lambda _k = \frac{1+\sqrt{1+4\lambda _{k-1}^2}}{2},~ \lambda _0 = 0. \end{aligned}$$\end{document} Here, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{g}^{(k)}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{g}\equiv \textbf{g}^{(\infty )}$$\end{document} are the inducing vectors of g and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$g_{\rho _{\gamma }}^*$$\end{document}, respectively. The complete AOMT algorithm is summarized in Algorithm 1. Algorithm 1Spherical area-measure-preserving OMT (AOMT) map construction with acceleration via the FISTA Spherical area-measure-preserving OMT (AOMT) map construction with acceleration via the FISTA Once the boundary \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\partial \mathcal {M}$$\end{document} is transformed to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\partial \mathcal {B}^3$$\end{document} by \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$g_{\rho _{\gamma }}^{*}$$\end{document} in (6), we set \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{f}^*_\texttt{B}= \textbf{g}$$\end{document} and define a homotopy \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$g_t: \partial \mathcal {M} \rightarrow \mathbb {R}^3$$\end{document} by\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} g_t(\textbf{v}) =& (1 - t) \textbf{v} + t g_{\rho _{\gamma }}^{*}(\textbf{v}),\\& \ \text{ for } \textbf{v} \in \mathbb {V}(\partial \mathcal {M}), \ 0 \le t \le 1. \end{aligned}$$\end{document}This homotopy is used to determine the interior points \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{f}^*_\texttt{I}$$\end{document} of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {B}^3$$\end{document} that minimizes the cost \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${C}_{\text{ v }}(f)$$\end{document} in (5). Let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ 0 = t_0< t_1< \cdots < t_p = 1$$\end{document}. According to (4), the interior map can be computed by solving the sequence of linear systems\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} &{[}L_V(f^{(k)})]_{\texttt{I},\texttt{I}} \textbf{f}^{(k+1)}_\texttt{I}\\&\quad= -[L_V(f^{(k)})]_{\texttt{I},\texttt{B}} \textbf{g}_{t_k},\ k = 0, 1, \ldots , p \end{aligned}$$\end{document}and we obtain \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{f}^*_\texttt{I}= \textbf{f}^{(p+1)}$$\end{document}. This method can be regarded as solving the nonlinear equation\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} {[}L_V(f)]_{\texttt{I},\texttt{I}} \textbf{f}_\texttt{I}= -[L_V(f)]_{\texttt{I},\texttt{B}} \textbf{f}_\texttt{B}, \end{aligned}$$\end{document}where each row of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{f}_\texttt{B}$$\end{document} has Euclidean norm equal to 1. Thus, the homotopy method can be viewed as a variation of the VSEM algorithm.Fig. 4A diagram illustrating the construction of the homotopy between \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(\mathcal {M}, \rho _{\gamma })$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(\mathcal {C}, \rho _{\gamma }^c \equiv 1)$$\end{document} A diagram illustrating the construction of the homotopy between \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(\mathcal {M}, \rho _{\gamma })$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(\mathcal {C}, \rho _{\gamma }^c \equiv 1)$$\end{document} By combining these two steps, we obtain a mass-preserving map \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f_{\rho _{\gamma }}^*$$\end{document} that transforms \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {M}$$\end{document} into \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {B}^3$$\end{document}, as illustrated in Fig. 3. Similarly, we can transform the solid cube \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {C}$$\end{document} with a constant density function \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rho _{\gamma }^c \equiv 1$$\end{document} into \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {B}^3$$\end{document} via the volume-preserving OMT map \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f_1^*: \mathcal {C} \rightarrow \partial \mathcal {B}^3$$\end{document}. Consequently, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{f}_{\rho _{\gamma }}^* = (f_1^*)^{-1} \circ f_{\rho _{\gamma }}^*$$\end{document} can be used to transform the irregular domain \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {M}$$\end{document} into the solid cube \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {C}$$\end{document} while preserving mass. However, due to the large number of vertices in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {M}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {C}$$\end{document}, directly computing \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(f_1^*)^{-1} \circ f_{\rho _{\gamma }}^*: \mathcal {M} \rightarrow \mathcal {C}$$\end{document} can be time consuming. To address this issue, we combine the AOMT maps \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$g_{\rho _{\gamma }}^*: \partial \mathcal {M} \rightarrow \partial \mathcal {B}^3$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$g_1^*: \partial \mathcal {C} \rightarrow \partial \mathcal {B}^3$$\end{document} to create the area-measure-preserving map \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{g}_{\rho _{\gamma }}^* = (g_1^*)^{-1} \circ g_{\rho _{\gamma }}^*: \partial \mathcal {M} \rightarrow \partial \mathcal {C}$$\end{document}. We then apply the homotopy method by solving\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} &{[}L_V(f^{(k)})]_{\texttt{I},\texttt{I}} \textbf{f}^{(k+1)}_{\texttt{I}}\\&\quad = -[L_V(f^{(k)})]_{\texttt{I},\texttt{B}} \textbf{g}_{c,t_k},\ k = 0, 1, \ldots , p \end{aligned}$$\end{document}with\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} g_{c,t}(\textbf{v}) =& (1 - t) \textbf{v} + t \hat{g}_{\rho _{\gamma }}^{*}(\textbf{v}), \\&\ \text{ for } \textbf{v} \in \mathbb {V}(\partial \mathcal {M}), \ 0 \le t \le 1. \end{aligned}$$\end{document}This allows us to compute the interior vertices of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {C}$$\end{document}, as illustrated in Fig. 4 and it is summarized in Algorithm 2. Algorithm 2Mass-preserving geometric transformation from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {M}$$\end{document} to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {C}$$\end{document} Mass-preserving geometric transformation from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {M}$$\end{document} to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {C}$$\end{document} We next explain the computation of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(g_1^*)^{-1} \circ g_{\rho _{\gamma }}^* (\partial \mathcal {M})$$\end{document} in Step 4 of Algorithm 2, as illustrated in Fig. 5. Let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{v} \in g_{\rho _{\gamma }}^*(\partial \mathcal {M}) \subset \partial \mathcal {B}^3$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f_1^*: \mathcal {C} \rightarrow \mathcal {B}^3$$\end{document} be the volume-preserving OMT map. Assuming that tetrahedron \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ [f_1^*(\textbf{v}_1), f_1^*(\textbf{v}_2), f_1^*(\textbf{v}_3), f_1^*(\textbf{v}_4)] \in \mathbb {T}(f_1^*(\mathcal {C})) $$\end{document} contains vertex \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ \textbf{v} $$\end{document} (where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ [\textbf{v}_1, \textbf{v}_2, \textbf{v}_3, \textbf{v}_4] \in \mathbb {T}(\mathcal {C}) $$\end{document}), we can express \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ \textbf{v} $$\end{document} as a linear combination of vertices \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ \{f_1^*(\textbf{v}_i)\}_{i=1}^4 $$\end{document} using barycentric coordinates as coefficients:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{v}&= \frac{|[\textbf{v}, f_1^*(\textbf{v}_2), f_1^*(\textbf{v}_3), f_1^*(\textbf{v}_4)]|}{|[f_1^*(\textbf{v}_1), f_1^*(\textbf{v}_2), f_1^*(\textbf{v}_3), f_1^*(\textbf{v}_4)]|}f_1^*(\textbf{v}_1) \\&\quad+ \frac{|[f_1^*(\textbf{v}_1), \textbf{v}, f_1^*(\textbf{v}_3), f_1^*(\textbf{v}_4)]|}{|[f_1^*(\textbf{v}_1), f_1^*(\textbf{v}_2), f_1^*(\textbf{v}_3), f_1^*(\textbf{v}_4)]|}f_1^*(\textbf{v}_2) \\&\quad+ \frac{|[f_1^*(\textbf{v}_1), f_1^*(\textbf{v}_2), \textbf{v}, f_1^*(\textbf{v}_4)]|}{|[f_1^*(\textbf{v}_1), f_1^*(\textbf{v}_2), f_1^*(\textbf{v}_3), f_1^*(\textbf{v}_4)]|}f_1^*(\textbf{v}_3) \\&\quad+ \frac{|[f_1^*(\textbf{v}_1), f_1^*(\textbf{v}_2), f_1^*(\textbf{v}_3), \textbf{v}]|}{|[f_1^*(\textbf{v}_1), f_1^*(\textbf{v}_2), f_1^*(\textbf{v}_3), f_1^*(\textbf{v}_4)]|}f_1^*(\textbf{v}_4) \\&\equiv \lambda _1 f_1^*(\textbf{v}_1) + \lambda _2 f_1^*(\textbf{v}_2) + \lambda _3 f_1^*(\textbf{v}_3) + \lambda _4 f_1^*(\textbf{v}_4). \end{aligned}$$\end{document}Since \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f_1^*$$\end{document} is piecewise linear and bijective, we have\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \hat{\textbf{v}} \equiv& (g_1^*)^{-1} (\textbf{v}) = (f_1^*)^{-1} (\textbf{v}) \\=&\, \lambda _1 \textbf{v}_1 + \lambda _2 \textbf{v}_2 + \lambda _3 \textbf{v}_3 + \lambda _4 \textbf{v}_4, \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{\textbf{v}}$$\end{document} is computed using the known \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda _i, \textbf{v}_i$$\end{document} for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$i = 1, \ldots , 4$$\end{document}.Fig. 5A diagram illustrating the computation of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(g_1^*)^{-1} \circ g_{\rho _{\gamma }}^*: \partial \mathcal {M} \rightarrow \partial \mathcal {C}$$\end{document} via the orange color flow A diagram illustrating the computation of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(g_1^*)^{-1} \circ g_{\rho _{\gamma }}^*: \partial \mathcal {M} \rightarrow \partial \mathcal {C}$$\end{document} via the orange color flow Two-phase segmentation with the ensemble technique for brain tumors In the previous section, we introduced the mass-preserving geometric approach, which transforms the irregular 3D brain domain \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {M}$$\end{document} with discrete tetrahedrons \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbb {T}(\mathcal {M})$$\end{document} and a specified density function \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rho _{\gamma }$$\end{document} into a solid cube \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {C}$$\end{document} with discrete tetrahedrons \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbb {T}(\mathcal {C})$$\end{document}. In this section, we discuss how to apply this mass-preserving geometric map to detect and segment brain tumors. Transformations of grayscale values and tumor labels To apply Algorithm 2 for brain tumor segmentation, we must first create an \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ m \times m \times m $$\end{document} MPGT tensor from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ \mathcal {C} $$\end{document}. This process involves two key transformations: (i) converting the grayscale values of the brain MRI data into the MPGT tensor and (ii) propagating the results predicted on the MPGT tensor back to the original brain MR image. Once these steps are addressed, the resulting tensors can be used as inputs for model training, enabling us to predict results for test data based on the MPGT tensor.Fig. 6A diagram illustrating the relationship between brain images and the MPGT tensor A diagram illustrating the relationship between brain images and the MPGT tensor Let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{f}_{\rho _{\gamma }}^*: \mathcal {M} \rightarrow \mathcal {C}$$\end{document} be the mass-preserving map generated via Algorithm 2. We denote \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbb {Q}(\mathcal {M})$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbb {Q}(\mathcal {C})$$\end{document} as the sets of all voxel centers for the brain MR image and the MPGT tensor, respectively (see Fig. 6). By applying \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(\hat{f}_{\rho _{\gamma }}^*)^{-1}$$\end{document} and using barycentric coordinates, we can transform the set \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbb {Q}(\mathcal {C})$$\end{document} into a new set \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbb {Q}(\mathcal {M}_c)$$\end{document} of brain query vertices. Establishing the relationship between the vertices in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbb {Q}(\mathcal {M}_c)$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbb {Q}(\mathcal {M})$$\end{document} allows us to convert the grayscale values in the MR image to the MPGT tensor and to map the labels in the MPGT tensor back to the MR image. For each \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{v} \in \mathbb {Q}(\mathcal {C})$$\end{document}, we identify a tetrahedron \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$[ \textbf{v}_1, \textbf{v}_2, \textbf{v}_3, \textbf{v}_4 ] \in \mathbb {T}(\mathcal {C})$$\end{document} that contains \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{v}$$\end{document}. We then express \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{v}$$\end{document} via barycentric coordinates as\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{v} = \lambda _1 \textbf{v}_1 + \lambda _2 \textbf{v}_2 + \lambda _3 \textbf{v}_3 + \lambda _4 \textbf{v}_4, \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$0 \le \lambda _i \le 1$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda _1 + \lambda _2 + \lambda _3 + \lambda _4 = 1$$\end{document}. This implies that\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \hat{\textbf{v}} \equiv &\,(\hat{f}_{\rho _{\gamma }}^*)^{-1}(\textbf{v}) \\=&\, \lambda _1 (\hat{f}_{\rho _{\gamma }}^*)^{-1}(\textbf{v}_1) + \lambda _2 (\hat{f}_{\rho _{\gamma }}^*)^{-1}(\textbf{v}_2) \\&+ \lambda _3 (\hat{f}_{\rho _{\gamma }}^*)^{-1}(\textbf{v}_3) + \lambda _4 (\hat{f}_{\rho _{\gamma }}^*)^{-1}(\textbf{v}_4), \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$[ (\hat{f}_{\rho _{\gamma }}^*)^{-1}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(\textbf{v}_1),$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ (\hat{f}_{\rho _{\gamma }}^*)^{-1}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(\textbf{v}_2), $$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(\hat{f}_{\rho _{\gamma }}^*)^{-1}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(\textbf{v}_3),$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ (\hat{f}_{\rho _{\gamma }}^*)^{-1}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(\textbf{v}_4) ] $$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\in \mathbb {T}(\mathcal {M})$$\end{document}. Since \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda _i$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(\hat{f}_{\rho _{\gamma }}^*)^{-1}(\textbf{v}_i)$$\end{document} for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$i = 1, \ldots , 4$$\end{document} are known, we can compute the vertex \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{\textbf{v}} \in \mathbb {Q}(\mathcal {M}_c)$$\end{document}. To determine the grayscale value of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{v} \in \mathbb {Q}(\mathcal {C})$$\end{document}, we find the closest \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{u} \in \mathbb {Q}(\mathcal {M})$$\end{document} to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{\textbf{v}}$$\end{document} and assign the grayscale value of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{u}$$\end{document} to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{v}$$\end{document}. In contrast, if we have predicted labels on \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbb {Q}(\mathcal {C})$$\end{document}, we identify the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{\textbf{v}}$$\end{document} closest to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{u}$$\end{document} and assign the label of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{v}$$\end{document} to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{u}$$\end{document}. Through this process, we can construct \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$m \times m \times m$$\end{document} MPGT tensors for the training data, which serve as inputs to train the model. We then infer the testing data within the corresponding MPGT tensors and finally convert the predicted results back to the raw MR images. Data preprocessing via algorithm 2 The OMT map with density function \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rho _{\gamma }$$\end{document} can be used to transform the irregular brain image into an \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$m \times m \times m$$\end{document} tensor, thereby maintaining the global information in the 3D MR image. To enhance the tumor region while preserving the volume of the non-tumor region in the OMT tensor, Lin et al. [28] proposed a method to define the density function \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rho _{\gamma }$$\end{document}. Let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbb {H}_{\text {e}}$$\end{document} denote the normalized contrast-enhanced histogram equalization grayscale image, and let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbb {H}_{\text {e}}(v)$$\end{document} represent its value at voxel v. Based on \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbb {H}_{\text {e}}$$\end{document}, we define a step-like density function as8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \rho ^s_\gamma (v) = {\left\{ \begin{array}{ll} \exp (\gamma \mathbb {H}_{\text {e}}(v)), & \text {if } v\in \mathcal {R}_{w}, \\ 1, & \text {otherwise}, \end{array}\right. } \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma $$\end{document} is a hyperparameter chosen from the interval [0, 2] and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {R}_{w}$$\end{document} denotes the WT region. We then use an image-filtering technique to convolve \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rho ^s_\gamma $$\end{document} with an \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\ell \times \ell \times \ell $$\end{document} box filter to define a smooth density function \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rho _\gamma $$\end{document}:9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} {\rho }_\gamma (v) = \rho ^s_\gamma (v)\otimes \frac{ {1}_{\ell \times \ell \times \ell }}{\ell ^3},\quad v\in \mathcal {M}. \end{aligned}$$\end{document}From the constraints in (1) and the definition in (8), we observe that \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$|f_{\rho _{\gamma }}^*(\tau )| > |\tau |$$\end{document} for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tau \in \mathcal {R}_{w}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$|f_{\rho _{\gamma }}^*(\tau )| = |\tau |$$\end{document} for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tau \notin \mathcal {R}_{w}$$\end{document}. This indicates that the tumor region is enlarged while the volume of the non-tumor region is preserved in the MPGT tensor. Furthermore, a higher value of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma $$\end{document} results in a larger tumor volume within the MPGT tensor, potentially increasing sensitivity but also distortion. Specifically, when \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma =0$$\end{document}, the transformation is purely volume-preserving. As \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma $$\end{document} increases, regions with higher histogram intensity (typically tumors) are expanded in the target cubic domain. We empirically limit \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma $$\end{document} to [0, 2] based on previous studies [28, 31], which showed that \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma > 2$$\end{document} leads to excessive distortion of non-tumor tissues, degrading the boundary definition of the tumor core. In Sect. 4.3, we perform a sensitivity analysis to justify this selection and examine its impact on segmentation performance. In the training data, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {R}_w$$\end{document} in (8) represents the labeled WT region. In this case, it is straightforward to define the density function \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rho _{\gamma }$$\end{document} in (9) to construct the MPGT tensors and subsequently train the model. However, the region \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {R}_w$$\end{document} is not available for the online validation or testing data. To address this challenge, Lin et al. [28] proposed a two-phase OMT algorithm to construct the associated density function. In Phase I of the two-phase OMT algorithm, the value \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rho _\gamma (v) = \exp (\gamma \mathbb {H}_{\text {e}}(v))$$\end{document} at each voxel v in the brain region is used to estimate the potential WT region \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {R}_w$$\end{document} for the online validation and testing data. Once this region \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {R}_w$$\end{document} is established, in Phase II, the MPGT tensors are constructed using \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {R}_w$$\end{document}, and the inference results are obtained from the trained model. This approach requires constructing two MPGT tensors with different density functions for each inference dataset. To streamline the inference process, we replace the MPGT tensors with the raw data as training inputs in Phase I, eliminating the need to construct MPGT tensors for the testing data in this phase. A flowchart illustrating the new two-phase nnU-Net method with the MPGT-based preprocessing approach is presented in Fig. 7. Importantly, nnU-Net method can be replaced with other training models, such as MedNeXt.Fig. 7A flow chart of the two-phase nnU-Net method with the MPGT-based preprocessing approach A flow chart of the two-phase nnU-Net method with the MPGT-based preprocessing approach The tumor region in the MPGT tensor can be enlarged by adjusting the parameter \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma $$\end{document} in (8). For our experiments, we set \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma = 1.0, 1.5, 1.75$$\end{document} and 2.0 to generate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$4n_t$$\end{document} augmented training MPGT tensors from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$n_t$$\end{document} training brain samples, thus increasing data diversity. We further apply standard data augmentation techniques to these \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$4n_t$$\end{document} tensors during training to prevent overfitting and improve the model’s robustness. Postprocessing In this subsection, we propose two postprocessing steps to enhance the inference results. Let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textrm{GT}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textrm{PD}$$\end{document} represent the ground truth and prediction, respectively, for the nested regions \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textrm{WT}\supset \textrm{TC}\supset \textrm{ET}$$\end{document}. Furthermore, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textrm{GT}^c$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textrm{PD}^c$$\end{document} denote the complementary sets of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textrm{GT}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textrm{PD}$$\end{document}, respectively. We consider the confusion matrix in Table 1.Table 1Confusion matrix for evaluating segmentation performance\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textrm{GT}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textrm{GT}^c$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textrm{PD}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {TP}$$\end{document} (true positive)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {FP}$$\end{document} (false positive)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textrm{PD}^c$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {FN}$$\end{document} (false negative)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {TN}$$\end{document} (true negative) Table 1Confusion matrix for evaluating segmentation performance\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textrm{GT}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textrm{GT}^c$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textrm{PD}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {TP}$$\end{document} (true positive)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {FP}$$\end{document} (false positive)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textrm{PD}^c$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {FN}$$\end{document} (false negative)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {TN}$$\end{document} (true negative) Confusion matrix for evaluating segmentation performance The Dice score [35] is defined as10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \text {Dice score}&= \frac{2|\text {GT} \cap \text {PD}|}{|\text {GT}|+|\text {PD}|}. \end{aligned}$$\end{document}Additionally, we introduce new metrics, including the lesion-wise Dice score and the Hausdorff distance 95 (HD95) measure [36], to evaluate the performance of the proposed method using the BraTS 2023 dataset. These metrics are defined as11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \text {Lesion-wise Dice score}&= \frac{\sum _{k = 1}^L \text {Dice}(l_k)}{\text {TP} + \text {FN} + \text {FP}},\nonumber \\ \text {Lesion-wise HD95}&= \frac{\sum _{k = 1}^L \text {HD95}(l_k)}{\text {TP} + \text {FN} + \text {FP}}. \end{aligned}$$\end{document}Here, L is the number of ground truth lesions calculated after dilation and let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\ell _k$$\end{document} be the k-th lesion, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$k = 1,\dots ,L$$\end{document}. In Subsection 3.2, we showed that our two-phase segmentation method yields two sets of predictions for each testing case: one from Phase I (raw data) and one from Phase II (MPGT tensors). Because these two stages are trained on data with different characteristics, we can combine (ensemble) their predictions to improve both voxel-wise Dice and lesion-wise Dice scores. Furthermore, comparing (10) and (11), we note that the denominator in (11) explicitly includes the term FP. This implies that a larger number of disconnected false-positive regions leads to a lower lesion-wise Dice score. Therefore, implementing a postprocessing step to remove small, disconnected false-positive regions is essential to enhance lesion-wise performance. To design these postprocessing steps, we analyze the voxels in the WT, TC, and ET regions in the 1251 training samples from the BraTS 2023 dataset to enhance the inference results on the online validation dataset. Let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$m_i$$\end{document} represent the total number of disconnected WT/TC/ET tumor lesions for the i-th training sample and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$v_{i,j}$$\end{document} denote the number of voxels in the j-th disconnected WT/TC/ET tumor lesion for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$j = 1, \ldots , m_i$$\end{document}. We define \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$v_i = \max \{ v_{i,j}; j = 1, \ldots , m_i\}$$\end{document}. The five smallest values of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$v_i$$\end{document} for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$i = 1, \ldots , 1251$$\end{document} are (1951, 2808, 3338, 3745, 4825) for the WT region, (0, 19, 157, 184, 325) for the TC region, and (0, 14, 19, 40, 52) for the ET region. These results indicate that some training samples contain only very small TC and ET tumors. The distributions of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$v_{i,j}$$\end{document} for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$j = 1, \ldots , m_i$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$i = 1, \ldots , 1251$$\end{document} are illustrated in Fig. 8. We observe that the percentages of lesions with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$v_{i,j} \ge 2000$$\end{document}, 1000, and 500 are 76.6%, 82.3%, and 80.0% for the WT, TC, and ET regions, respectively.Fig. 8Analysis of the tumor voxels for each disconnected WT, TC, and ET region by using the distribution of the number (y-axis) of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$v_{i,j}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$j = 1, \ldots , m_i$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$i = 1, \ldots , 1251$$\end{document} belonging to the given interval (x-axis) Analysis of the tumor voxels for each disconnected WT, TC, and ET region by using the distribution of the number (y-axis) of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$v_{i,j}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$j = 1, \ldots , m_i$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$i = 1, \ldots , 1251$$\end{document} belonging to the given interval (x-axis) To address the observed difference between the Dice and lesion-wise Dice scores in (10)-(11), we establish volumetric thresholds for the WT, TC, and ET regions. If the number of voxels in a disconnected WT/TC/ET region in the inference data is below the specified threshold, we classify this region as a false positive and relabel it as a non-tumor region. Importantly, in the BraTS 2023 challenge, a volumetric threshold of 50 voxels is used, and lesions smaller than this threshold are not evaluated as true lesions [36]. Our analysis of the training data (Fig. 8) shows that there are only 32 disconnected WT lesions with tumor voxels in the interval (50, 250], 50 TC lesions with tumor voxels in (50, 450], and 90 ET lesions with tumor voxels in (50, 1450]. In addition, there are 23 disconnected TC lesions with tumor voxels in (50, 150] and 30 disconnected ET lesions with tumor voxels in (50, 100]. Thus, it is reasonable to set the volumetric thresholds to (450, 150, 80) voxels for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(\text {WT}, \text {TC}, \text {ET})$$\end{document}, respectively. In Subsection 4.5, we demonstrate the improvement in the lesion-wise Dice scores obtained after applying these thresholds. Numerical results The BraTS 2023 Adult Glioma Challenge dataset includes a total of 2040 brain images, with 1251 training images and 219 unlabeled samples for online validation. The remaining images are withheld for testing purposes. We set \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\ell = 5$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma \in \{1, 1.5, 1.75, 2\}$$\end{document} in (9) to construct \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$128 \times 128 \times 128$$\end{document} MPGT tensors for the 1251 training images, resulting in 5004 training MPGT tensors. For online validation, we used \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma \in \{ 1, 1.5, 1.75\}$$\end{document}. We used SegResNet, nnU-Net, and MedNeXt to train our models. SegResNet and MedNeXt were implemented in PyTorch using the MONAI framework, whereas nnU-Net was obtained from its open-source GitHub repository. All training sessions were conducted on a Linux server equipped with two NVIDIA Tesla A100 PCIe 40 GB GPUs.Fig. 9Convergence behavior of Algorithm 2 with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p = 9$$\end{document} Convergence behavior of Algorithm 2 with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p = 9$$\end{document} Mass-preserving property and bijectivity of the MPGT in algorithm 2 We utilize Algorithm 2 to compute the cubic mass-preserving mapping from the training data to the solid cube \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {C}$$\end{document}. The computed results f at various steps t of the homotopy method with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p = 9$$\end{document} (see Fig. 9) demonstrate that the raw data are progressively transformed into a solid cube, and the transformation is completed when \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t = 1$$\end{document}. Consequently, we need to solve only p linear systems to efficiently compute the solid cube \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {C}$$\end{document} once \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{g}_{\rho _{\gamma }}^*$$\end{document} is provided. Following [32], we introduce the local mass ratio to assess the accuracy of the cubic mass-preserving map. The local mass ratio \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {R}_f(\textbf{v})$$\end{document} at vertex \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{v}$$\end{document} is defined as\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \mathcal {R}_f(\textbf{v}) = \frac{\sum _{\tau \in N(\textbf{v})} \rho _{\gamma }(\tau ) | \tau |/\sum _{\tau \in \mathbb {T}(\mathcal {M})} \rho _{\gamma }(\tau ) | \tau |}{\sum _{\tau \in N(\textbf{v})} | f(\tau ) |/ | f(\mathcal {M}) |}. \end{aligned}$$\end{document}We can numerically assess the bijectivity of the cubic mass-preserving mapping by counting the number of folded tetrahedrons within the solid cube \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {C}$$\end{document}; the procedure is described in detail in [32]. Let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mu _i = \sum _{\textbf{v} \in \mathbb {T}(\mathcal {M}_i)} \mathcal {R}_f(\textbf{v}) / m_i$$\end{document} denote the mean value of all local mass ratios \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {R}_f(\textbf{v})$$\end{document} for the i-th training dataset \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {M}_i$$\end{document}, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$m_i$$\end{document} is the number of vertices in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {M}_i$$\end{document}. Figure 10a and 10b show box plots of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mu _i$$\end{document} and the corresponding folding numbers for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$i = 1, \ldots , 1251$$\end{document} with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma \in \{ 1, 1.5, 1.75, 2 \}$$\end{document}. The means of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mu _i$$\end{document} and the folding numbers are summarized in Table 2.Fig. 10Analysis of the mass ratio and folding number for Algorithm 2 with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma = 1, 1.5, 1.75, 2.0$$\end{document}Table 2Mean value and standard deviation (std) of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mu _i$$\end{document} and the folding number for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$i = 1, \ldots , 1251$$\end{document} vs. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma .$$\end{document}Mass ratioFolding number\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma $$\end{document}1.01.51.752.01.01.51.752.0Mean1.00171.00030.99850.99556.3710.5414.0619.59STD0.00410.01060.01570.02225.977.879.0211.37 Analysis of the mass ratio and folding number for Algorithm 2 with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma = 1, 1.5, 1.75, 2.0$$\end{document} Table 2Mean value and standard deviation (std) of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mu _i$$\end{document} and the folding number for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$i = 1, \ldots , 1251$$\end{document} vs. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma .$$\end{document}Mass ratioFolding number\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma $$\end{document}1.01.51.752.01.01.51.752.0Mean1.00171.00030.99850.99556.3710.5414.0619.59STD0.00410.01060.01570.02225.977.879.0211.37 Mean value and standard deviation (std) of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mu _i$$\end{document} and the folding number for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$i = 1, \ldots , 1251$$\end{document} vs. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma .$$\end{document} The means of the local mass ratios are close to 1, with standard deviations less than 0.02, indicating that the resulting MPGT maps have the satisfactory mass-preserving properties. Furthermore, the average number of tetrahedrons in the training datasets \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {M}_i$$\end{document} (for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$i = 1, \ldots , 1251$$\end{document}) is 635610, with a standard deviation of 22093. According to Table 2, the average percentage of folded tetrahedrons is less than \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$19.59/635610 \approx 0.003\%$$\end{document}. Therefore, the resulting MPGT maps are numerically very close to bijective. Computational efficiency comparison To validate the efficiency of our proposed homotopy-based MPGT in Algorithm 2, we compare it with the traditional volumetric OMT method [30]. The speedup for each BraTS validation case is reported in Fig. 11.Fig. 11Computational efficiency of the proposed homotopy-based MPGT (Algorithm 2) relative to the traditional volumetric OMT method, illustrating speedup per validation case Computational efficiency of the proposed homotopy-based MPGT (Algorithm 2) relative to the traditional volumetric OMT method, illustrating speedup per validation case On average, our method achieves a \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.31\times $$\end{document} speedup. This improvement arises because computing \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(f^{*}_{1})^{-1} \circ f^{*}_{\rho _\gamma }: \mathcal {M} \rightarrow \mathcal {C}$$\end{document} in the volumetric OMT method is time consuming for the large meshes involved (more than \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$10^5$$\end{document} vertices in both \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {M}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {C}$$\end{document}; see Subsection 2.3). In contrast, our homotopy-based approach transforms the interior vertices of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {M}$$\end{document} into the interior of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {C}$$\end{document} more efficiently by restricting the expensive computations to the spherical boundary and propagating the deformation inward. This makes our geometric transformation more practical for large 3D medical images. In terms of wall-clock time, the average runtimes per case for (i) constructing the MPGT tensor, (ii) inference in Phase I, (iii) inference in Phase II, and (iv) inverting the labels from the MPGT tensor back to the raw MRI space and ensembling the final prediction are 87.08, 8.17, 4.37, and 1.48 s, respectively. Thus, the total average time required by our complete pipeline to obtain a final inference result per case is approximately 101.11 s. Ablation study of the two-phase strategy and hyperparameter sensitivity This subsection presents the inference results of nnU-Net, MedNeXt, and SegResNet on the online validation set. An averaging ensemble, a common strategy in medical image analysis, was applied across five folds of nnU-Net to generate the predictions. Table 3 reports the Dice and lesion-wise Dice scores for Phases I and II for nnU-Net, MedNeXt, and SegResNet. The entry “Phase II (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma = 1$$\end{document})” in the table indicates that \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma = 1$$\end{document} was used to construct the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$128 \times 128 \times 128$$\end{document} MPGT tensors for online validation. The results in Table 3 demonstrate that the mean Dice scores and lesion-wise Dice scores in Phase II are generally higher than those in Phase I, except for the mean lesion-wise Dice score for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma = 1.75$$\end{document} in nnU-Net.Table 3Dice score and lesion-wise Dice score for Phases I and II with various \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma $$\end{document} and training modelsDice score \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\uparrow $$\end{document}Lesion-wise dice score \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\uparrow $$\end{document}MethodWTTCETMeanWTTCETMeannnU-Net [13]Phase I0.92700.87980.83800.88160.82960.85420.80430.8294Phase II (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma = 1.0$$\end{document})0.92500.88320.84530.88450.86700.85940.80800.8448Phase I + II0.92810.88010.84860.88560.86050.85490.81440.8433Phase II (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma = 1.5$$\end{document})0.92530.88290.84520.88450.85400.85390.80440.8374Phase I + II0.92810.88050.85360.88740.85390.85930.82040.8445Phase II (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma = 1.75$$\end{document})0.92500.88260.85000.88590.84820.84080.79620.8284Phase I + II0.92810.88110.84890.88600.85300.85390.81300.8400MedNeXt [14]Phase I0.92640.87370.84570.88190.81790.84480.81790.8267Phase II (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma = 1.75$$\end{document})0.92550.87760.84790.88370.85720.83670.80000.8313Phase I + II0.92740.87690.84950.88460.85870.84590.81660.8404SegResNet [12]Phase I0.91330.86540.77970.85280.68250.78680.68750.7189Phase II (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma = 1.5$$\end{document})0.92210.86730.84100.87680.67030.79510.78540.7503Phase I + II0.92180.87200.83630.87670.73440.81800.77970.7774 Table 3Dice score and lesion-wise Dice score for Phases I and II with various \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma $$\end{document} and training modelsDice score \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\uparrow $$\end{document}Lesion-wise dice score \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\uparrow $$\end{document}MethodWTTCETMeanWTTCETMeannnU-Net [13]Phase I0.92700.87980.83800.88160.82960.85420.80430.8294Phase II (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma = 1.0$$\end{document})0.92500.88320.84530.88450.86700.85940.80800.8448Phase I + II0.92810.88010.84860.88560.86050.85490.81440.8433Phase II (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma = 1.5$$\end{document})0.92530.88290.84520.88450.85400.85390.80440.8374Phase I + II0.92810.88050.85360.88740.85390.85930.82040.8445Phase II (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma = 1.75$$\end{document})0.92500.88260.85000.88590.84820.84080.79620.8284Phase I + II0.92810.88110.84890.88600.85300.85390.81300.8400MedNeXt [14]Phase I0.92640.87370.84570.88190.81790.84480.81790.8267Phase II (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma = 1.75$$\end{document})0.92550.87760.84790.88370.85720.83670.80000.8313Phase I + II0.92740.87690.84950.88460.85870.84590.81660.8404SegResNet [12]Phase I0.91330.86540.77970.85280.68250.78680.68750.7189Phase II (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma = 1.5$$\end{document})0.92210.86730.84100.87680.67030.79510.78540.7503Phase I + II0.92180.87200.83630.87670.73440.81800.77970.7774 Dice score and lesion-wise Dice score for Phases I and II with various \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma $$\end{document} and training models To assess the necessity of the two-phase pipeline, we evaluated the performance of Phase I (raw data) and Phase II (MPGT tensors) independently before combination. As shown in Table 3, Phase I (processing raw data directly) provides a strong baseline for global tumor localization. Phase II (utilizing MPGT tensors) alone yields high sensitivity but benefits from the more accurate localization provided by Phase I. The combination (Phase I + II) consistently achieves the highest mean Dice and lesion-wise Dice scores across all tumor sub-regions for nnU-Net, MedNeXt, and SegResNet. This demonstrates that the MPGT tensors provide complementary structural information that raw data alone may miss, especially for nnU-Net where the combined approach improve the lesion-wise Dices score for the WT by more 2.3% compared with Phase I alone. We further analyzed the model’s sensitivity to the hyperparameter \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma $$\end{document}. As indicated in Table 3, performance remains stable for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma \in [1.0, 1.75]$$\end{document}. However, as shown in Table 1, values of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma > 2$$\end{document} lead to excessive distortion of non-tumor tissues, which slightly degrades the boundary definition of the tumor core, whereas \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma < 1.0$$\end{document} fails to sufficiently magnify small lesions. Consequently, to ensure robustness and reproducibility across varying tumor sizes, we employ data augmentation with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma \in \{1, 1.5, 1.75, 2 \}$$\end{document} for the training set and select \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma \in \{1, 1.5, 1.75 \}$$\end{document} for the validation set. Statistical significance analysis Table 4Statistical significance analysis comparing the baseline model (Phase I) with the Phase II model and the Phase I + II ensemble model on the validation dataset. The reported p-values are computed using the Wilcoxon signed-rank test across validation casesPhase I vs Phase IIPhase I vs Phase I + IIMetricWTTCETWTTCETDice score (p-value)0.00020.00520.00050.00190.00930.0033Lesion-wise dice (p-value)0.09440.00030.00040.00010.03700.0014 Table 4Statistical significance analysis comparing the baseline model (Phase I) with the Phase II model and the Phase I + II ensemble model on the validation dataset. The reported p-values are computed using the Wilcoxon signed-rank test across validation casesPhase I vs Phase IIPhase I vs Phase I + IIMetricWTTCETWTTCETDice score (p-value)0.00020.00520.00050.00190.00930.0033Lesion-wise dice (p-value)0.09440.00030.00040.00010.03700.0014 Statistical significance analysis comparing the baseline model (Phase I) with the Phase II model and the Phase I + II ensemble model on the validation dataset. The reported p-values are computed using the Wilcoxon signed-rank test across validation cases Although Table 3 shows consistent improvements in Dice scores and lesion-wise Dice scores, it is important to verify whether these performance gains are statistically significant. Therefore, we performed a statistical significance analysis on the validation dataset. Since the baseline model (Phase I) and the compared methods were evaluated on the same cases, paired statistical testing was applied. Specifically, the nonparametric Wilcoxon signed-rank test was used to compare the performance of Phase II and the Phase I + II ensemble against the baseline Phase I model for the three tumor subregions: WT, TC, and ET. The resulting p-values are summarized in Table 4. Most of the p-values are below 0.05, indicating that the improvements achieved by Phase II and the Phase I + II ensemble over the baseline Phase I model are statistically significant for the majority of evaluated metrics. Effect of the threshold The Dice and lesion-wise Dice scores in each row of Table 3 show noticeable differences. As indicated in (10)-(11), these discrepancies suggest the presence of false negative or false positive lesions. To mitigate the impact of small false positives and improve lesion-wise Dice scores, we apply volumetric thresholds after ensembling Phase I and Phase II results, thereby filtering out small disconnected lesions. On the basis of the analysis in Subsection 3.3, we use thresholds of (450, 150, 80) voxels for (WT, TC, ET) for nnU-Net, MedNeXt, and SegResNet. Furthermore, Ferreira et al. [23] used thresholds of 250 and 1450 voxels for the WT region. To facilitate comparison with [23], we also apply thresholds of (250, 150, 80) and (1450, 150, 80) for nnU-Net. The corresponding Dice and lesion-wise Dice scores for the WT, TC, and ET regions at these thresholds are presented in Table 5. The threshold (0, 0, 0) represents the original ensemble of Phase I and Phase II results without thresholding, corresponding to the “Phase I + II” results in Table 3. Comparing the results in Table 5, we observe only minor changes in the Dice scores for the WT, TC, and ET regions with and without thresholds. However, the lesion-wise Dice scores for the WT, TC, and ET regions significantly improve after applying the thresholds. Consequently, the discrepancies between Dice and lesion-wise Dice scores can be substantially reduced by using appropriate thresholds.Table 5Ensemble results for Phases I and II under different volumetric thresholdsThresholds (WT, TC, ET)Dice score \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\uparrow $$\end{document}Lesion-wise dice score \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\uparrow $$\end{document}WTTCETMeanWTTCETMeannnU-Net, Phase I + II (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma = 1.0$$\end{document})(0, 0, 0)0.92810.88010.84860.88560.86050.85490.81440.8433(250, 150, 80)0.92810.88020.84490.88440.90380.86590.84310.8709(450, 150, 80)0.92810.88020.84490.88440.90010.86590.84310.8697(1450, 150, 80)0.92810.88020.84490.88440.90710.86590.84310.8720nnU-Net, Phase I + II (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma = 1.5$$\end{document})(0, 0, 0)0.92810.88050.85360.88740.85390.85930.82040.8445(250, 150, 80)0.92820.88030.85290.88710.90390.86830.85590.8760(450, 150, 80)0.92820.88030.85290.88710.90450.86830.85590.8762(1450, 150, 80)0.92810.88030.85290.88710.90710.86830.85590.8771nnU-Net, Phase I + II (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma = 1.75$$\end{document})(0, 0, 0)0.92810.88110.84890.88600.85300.85390.81300.8400(250, 150, 80)0.92810.88120.85270.88730.90380.86920.85580.8763(450, 150, 80)0.92820.88120.85270.88740.90430.86920.85580.8764(1450, 150, 80)0.92810.88110.85270.88730.90700.86920.85580.8773MedNeXt, Phase I + II (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma = 1.75$$\end{document})(0, 0, 0)0.92740.87690.84950.88460.85870.84590.81660.8404(450, 150, 80)0.92740.87650.84730.88370.90160.86520.84790.8716SegResNet, Phase I + II (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma = 1.5$$\end{document})(0, 0, 0)0.92180.87200.83630.87670.73440.81800.77970.7774(450, 150, 80)0.92180.87600.83980.87920.89400.86430.84480.8677 Table 5Ensemble results for Phases I and II under different volumetric thresholdsThresholds (WT, TC, ET)Dice score \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\uparrow $$\end{document}Lesion-wise dice score \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\uparrow $$\end{document}WTTCETMeanWTTCETMeannnU-Net, Phase I + II (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma = 1.0$$\end{document})(0, 0, 0)0.92810.88010.84860.88560.86050.85490.81440.8433(250, 150, 80)0.92810.88020.84490.88440.90380.86590.84310.8709(450, 150, 80)0.92810.88020.84490.88440.90010.86590.84310.8697(1450, 150, 80)0.92810.88020.84490.88440.90710.86590.84310.8720nnU-Net, Phase I + II (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma = 1.5$$\end{document})(0, 0, 0)0.92810.88050.85360.88740.85390.85930.82040.8445(250, 150, 80)0.92820.88030.85290.88710.90390.86830.85590.8760(450, 150, 80)0.92820.88030.85290.88710.90450.86830.85590.8762(1450, 150, 80)0.92810.88030.85290.88710.90710.86830.85590.8771nnU-Net, Phase I + II (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma = 1.75$$\end{document})(0, 0, 0)0.92810.88110.84890.88600.85300.85390.81300.8400(250, 150, 80)0.92810.88120.85270.88730.90380.86920.85580.8763(450, 150, 80)0.92820.88120.85270.88740.90430.86920.85580.8764(1450, 150, 80)0.92810.88110.85270.88730.90700.86920.85580.8773MedNeXt, Phase I + II (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma = 1.75$$\end{document})(0, 0, 0)0.92740.87690.84950.88460.85870.84590.81660.8404(450, 150, 80)0.92740.87650.84730.88370.90160.86520.84790.8716SegResNet, Phase I + II (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma = 1.5$$\end{document})(0, 0, 0)0.92180.87200.83630.87670.73440.81800.77970.7774(450, 150, 80)0.92180.87600.83980.87920.89400.86430.84480.8677 Ensemble results for Phases I and II under different volumetric thresholds The results in Table 5 indicate that, for nnU-Net, the lesion-wise Dice score for the WT region increases as the WT threshold increases. The threshold of 1450 voxels for WT yields the highest lesion-wise Dice score among the tested thresholds. However, as illustrated in Fig. 12, the small disconnected WT lesion in the right brain image is removed when the 1450-voxel threshold is used, which causes a substantial decrease in the lesion-wise Dice score from 0.8132 to 0.5025 for that case. Therefore, a threshold of 1450 voxels for the WT region may be too high for practical applications [23]. In contrast, a threshold of 450 voxels for the WT region appears more suitable for data inference.Fig. 12The left and right brain images with the predicted WT regions, showing the results with and without the threshold of 1450 for ID BraTS-GLI-00699-001. The associated lesion-wise Dice scores of the WT are 0.5025 and 0.8132 The left and right brain images with the predicted WT regions, showing the results with and without the threshold of 1450 for ID BraTS-GLI-00699-001. The associated lesion-wise Dice scores of the WT are 0.5025 and 0.8132 Comparison results with various methods As demonstrated in Subsections 4.3 and 4.5, our proposed ensemble strategy and the volumetric thresholds for WT, TC, and ET significantly improve the lesion-wise Dice scores for SegResNet, nnU-Net, and MedNeXt. Among these three models, nnU-Net achieves the highest Dice and lesion-wise Dice scores for the WT, TC, and ET regions. In this subsection, we compare our nnU-Net results with those of the winner [23] and the third place model [25] from the BraTS 2023 Adult Glioma Segmentation Challenge. We also compare our Dice scores with those of winners from 2021 to 2023. The winning team [23] employed three models: the baseline nnU-Net (B) [16], Swin UNETR (S) [24], and the BraTS 2021 winning model (L) [15]. They also used three methods for generating synthetic data, denoted as G, rG, and R. Rows 4-7 of Table 6 present the lesion-wise Dice scores and HD95 results reported in [23]. In Table 6, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S_B^G$$\end{document} refers to the solution obtained using the baseline network B with synthetic data G, while \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S_{B,S}^{G,rG}$$\end{document} represents the ensemble of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S_B^G$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S_S^{rG}$$\end{document}. The entry “All” corresponds to the combined ensemble \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S_{B,L,S,B,L,S,B,L,S}^{G,G,G,rG,rG,rG,R,R,R}$$\end{document}.Table 6Lesion-wise dice scores and lesion-wise Hausdorff distance (HD95) valuesLesion-wise dice score \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\uparrow $$\end{document}Lesion-wise HD95 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\downarrow $$\end{document}MethodWTTCETMeanWTTCETMeanWinner [23]All (1450, 150, 100)0.91010.86730.85090.876111.1114.4717.7014.43All (250, 150, 100)0.90050.86730.85090.872914.9414.4717.7015.70\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S^{rG,rG,rG}_{B,L,S}$$\end{document} (250, 150, 100)0.89720.86860.85270.872815.5114.4217.6415.86\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S^{G,G,G}_{B,L,S}$$\end{document} (300, 200, 200)0.89690.86600.85280.871915.8016.4019.4717.22Third place [25]MSS0.90630.86270.84320.870711.7013.1017.3714.06MS0.90630.86220.83460.867711.7118.7123.7018.04Ours (nnU-Net, Phase I + II, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma = 1.75$$\end{document})(250, 150, 80)0.90380.86920.85580.876310.7412.0814.0712.86(450, 150, 80)0.90430.86920.85580.876412.5712.0814.0712.91(1450, 150, 80)0.90700.86920.85580.877311.7912.0914.0712.65 Table 6Lesion-wise dice scores and lesion-wise Hausdorff distance (HD95) valuesLesion-wise dice score \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\uparrow $$\end{document}Lesion-wise HD95 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\downarrow $$\end{document}MethodWTTCETMeanWTTCETMeanWinner [23]All (1450, 150, 100)0.91010.86730.85090.876111.1114.4717.7014.43All (250, 150, 100)0.90050.86730.85090.872914.9414.4717.7015.70\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S^{rG,rG,rG}_{B,L,S}$$\end{document} (250, 150, 100)0.89720.86860.85270.872815.5114.4217.6415.86\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S^{G,G,G}_{B,L,S}$$\end{document} (300, 200, 200)0.89690.86600.85280.871915.8016.4019.4717.22Third place [25]MSS0.90630.86270.84320.870711.7013.1017.3714.06MS0.90630.86220.83460.867711.7118.7123.7018.04Ours (nnU-Net, Phase I + II, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma = 1.75$$\end{document})(250, 150, 80)0.90380.86920.85580.876310.7412.0814.0712.86(450, 150, 80)0.90430.86920.85580.876412.5712.0814.0712.91(1450, 150, 80)0.90700.86920.85580.877311.7912.0914.0712.65 Lesion-wise dice scores and lesion-wise Hausdorff distance (HD95) values The third-place team [25] employed SegResNet and MedNeXt to segment three tumor subregions and introduced several strategies to better detect small lesions. Their best result, denoted by MSS (in the ninth row of Table 6), includes an ensemble of MedNeXt-DS, SegResNet-DS, and SegResNet; MS (in the tenth row of Table 6) denotes the ensemble of MedNeXt-DS and SegResNet-DS, where “DS” refers to training with deep supervision. Rows 12 to 14 of Table 6 report the results of our method using nnU-Net with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma = 1.75$$\end{document}. Excluding the rows obtained with a WT threshold of 1450 voxels, our results for the threshold (450, 150, 80) show that only the lesion-wise Dice score and HD95 for the WT region are slightly worse than those of the third-place model. All other lesion-wise Dice and HD95 results are better than those of both the winner and the third-place method. Furthermore, the mean lesion-wise Dice and HD95 values over WT, TC, and ET are higher and lower, respectively, than those reported by both teams.Table 7Dice scores and Hausdorff distance (HD95) values for BraTS 2021–2023 winners and our methodDice score \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\uparrow $$\end{document}HD95 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\downarrow $$\end{document}MethodWTTCETMeanWTTCETMean2023 winner (All) [23]0.92940.87690.84640.88424.2611.1217.8111.062023 winner (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S^{rG,rG,rG}_{B,L,S}$$\end{document})0.92860.87810.84840.88504.1911.0817.7511.002023 winner (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S^{G,G,G}_{B,L,S}$$\end{document})0.92830.87610.85150.88534.2712.7019.3212.102022 winner [17]0.92710.87530.84380.88213.607.5317.509.542021 winner [15]0.92750.87810.84510.88363.477.6220.7310.61Ours (nnU-Net, Phase I + II, threshold (450, 150, 80))\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma = 1.5$$\end{document}0.92820.88030.85290.88714.449.5412.968.98\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma = 1.75$$\end{document}0.92820.88120.85270.88744.447.8512.968.42 Table 7Dice scores and Hausdorff distance (HD95) values for BraTS 2021–2023 winners and our methodDice score \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\uparrow $$\end{document}HD95 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\downarrow $$\end{document}MethodWTTCETMeanWTTCETMean2023 winner (All) [23]0.92940.87690.84640.88424.2611.1217.8111.062023 winner (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S^{rG,rG,rG}_{B,L,S}$$\end{document})0.92860.87810.84840.88504.1911.0817.7511.002023 winner (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S^{G,G,G}_{B,L,S}$$\end{document})0.92830.87610.85150.88534.2712.7019.3212.102022 winner [17]0.92710.87530.84380.88213.607.5317.509.542021 winner [15]0.92750.87810.84510.88363.477.6220.7310.61Ours (nnU-Net, Phase I + II, threshold (450, 150, 80))\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma = 1.5$$\end{document}0.92820.88030.85290.88714.449.5412.968.98\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma = 1.75$$\end{document}0.92820.88120.85270.88744.447.8512.968.42 Dice scores and Hausdorff distance (HD95) values for BraTS 2021–2023 winners and our method In [23], the authors used generative adversarial networks and registration techniques to generate 23049 training samples for their three deep-learning models. This data generation process took approximately two weeks. In contrast, our method generates only 5004 training MPGT tensors in Phase II using four different values of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma $$\end{document}, while still preserving the global MRI features in each tensor. Moreover, the “All" ensemble in [23] combines nine models (three architectures with three synthetic data schemes), each trained via five-fold cross-validation, leading to an ensemble of 45 inference passes for the final prediction. In our case, we ensemble only ten inference results (5 folds \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times $$\end{document} 2 phases) using nnU-Net. Despite using less training data and fewer inference passes than the winner in [23], our results in Table 6 show that our method achieves superior lesion-wise Dice scores and HD95 metrics, except the lesion-wise Dice score for WT at the thresholds (1450, 150, 100). These findings highlight the effectiveness of the MPGT representation with hyperparameter \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma $$\end{document} and our ensemble strategy. The same dataset were used for training and online validation in BraTS 2021, BraTS 2022, and BraTS 2023. Table 7 reports the Dice scores and HD95 values for the WT, TC, and ET regions for the winners of these years and for our proposed method. The results in Tables 6 and 7 indicate that our method consistently outperforms, or is competitive with, the approaches in [15, 17, 23, 25] in terms of both Dice and lesion-wise Dice scores. Conclusion The two-stage optimal mass transport (OMT) method with a mass-preserving map was proposed by Lin et al. [27] to transform irregular brain MR images into cubes, with the generated 3D tensor used as the input to a 3D U-Net model. Building on this OMT framework, Lin et al. [28] introduced a two-phase residual U-Net algorithm. In Phase I, MPGT tensors are used to accurately predict the WT region, whereas in Phase II, a new MPGT tensor is constructed based on this prediction, requiring two MPGT tensors per case. In this article, we propose a modified mass-preserving geometric transformation that uses a homotopy method to reduce computational cost in the two-stage process. We further develop a modified two-phase segmentation approach to minimize inference time for testing data. By replacing MPGT tensors with randomly cropped raw data in Phase I, we only need to compute MPGT tensors in Phase II. We also introduce postprocessing steps that ensemble the Phase I and Phase II results and apply volumetric thresholds for tumor detection to improve lesion-wise performance on the testing data. Our numerical results show that using SegResNet, nnU-Net, and MedNeXt within the proposed framework yields comparable Dice scores and lesion-wise Dice scores to state-of-the-art models during online validation, demonstrating the robustness of the proposed approach. Among these models, nnU-Net achieves the highest Dice scores (0.9282, 0.8812, 0.8527) and lesion-wise Dice scores (0.9043, 0.8692, 0.8558), outperforming or showing comparable performance to the models in [15, 17, 23, 25]. Regarding generalizability and external validation, we acknowledge that, although our method demonstrates high performance on the well-curated BraTS 2023 dataset, the lack of validation on independent external cohorts is a current limitation for assessing robustness in real-world scenarios. The BraTS dataset is highly curated, whereas real-world clinical data often contains artifacts, varying resolutions, and diverse scanner protocols that are not fully represented here, which can lead to significant domain-shift issues. The hyperparameter \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma $$\end{document}, used to enlarge tumor regions in the MPGT space, was empirically optimized for gliomas, and its applicability to other tumor types (e.g., meningiomas or metastases) requires further validation. Furthermore, the postprocessing volume thresholds (450, 150, 80) voxels were derived from the training set statistics. Therefore, future work will focus on adaptive thresholding techniques and extensive cross-dataset validation on external cohorts (e.g., the Erasmus Glioma database (EGD) dataset [37], the LUMIERE dataset [38], the University of Pennsylvania glioblastoma (UPenn-GBM) dataset [39], The Cancer Genome Atlas Glioblastoma Multiforme (TCGA-GBM) dataset, and the Cancer Genome Atlas Low Grade Glioma (TCGA-LGG) dataset [40]) to ensure robustness against such domain shifts and to further improve the method’s generalizability.