Anatomy-DT: A Cross-Diffusion Digital Twin for Anatomical Evolution
Moinak Bhattacharya, Gagandeep Singh & Prateek Prasanna · 2025
Read Anatomy-DT: A Cross-Diffusion Digital Twin for Anatomical Evolution by Moinak Bhattacharya, Gagandeep Singh & Prateek Prasanna (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2509-25280.txt.
Introduction
Moinak Bhattacharya, Gagandeep Singh, and Prateek Prasanna build a digital twin that grows a tumor together with the tissue around it. Most models of disease progression track the tumor alone, yet real tumors evolve unevenly, shaped by treatment and by the tissues they push against. A twin that ignores surrounding anatomy misses half the clinical story.
Their framework, Anatomy-DT, couples physics-based equations with trainable deep learning. Anatomy is encoded as a field of class probabilities and evolved by a cross-diffusion system that enforces competition between tissue types, keeping each region purely one type. A topology penalty keeps centerlines intact and punishes overlaps between regions. On synthetic benchmarks the method reached top accuracy while preserving anatomy, and it also led on a clinical dataset. The authors present the system as a principled route to realistic simulation of whole anatomy that stays consistent.
Catalog record
| CREATOR | Moinak Bhattacharya, Gagandeep Singh & Prateek Prasanna |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 39 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2509-25280
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Anatomy-DT: A Cross-Diffusion Digital Twin for Anatomical Evolution Moinak Bhattacharya, Gagandeep Singh & Prateek Prasanna Abstract Accurately modeling the spatiotemporal evolution of tumor morphology from baseline imaging is a pre-requisite for developing digital twin frameworks that can simulate disease progression and treatment response. Most existing approaches primarily characterize tumor growth while neglecting the concomitant alterations in adjacent anatomical structures. In reality, tumor evolution is highly non-linear and heterogeneous, shaped not only by therapeutic interventions but also by its spatial context and interaction with neighboring tissues. Therefore, it is critical to model tumor progression in conjunction with surrounding anatomy to obtain a comprehensive and clinically relevant understanding of disease dynamics. We introduce a mathematically grounded framework that unites mechanistic partial differential equations (PDEs) with differentiable deep learning. Anatomy is represented as a multi-class probability field on the simplex and evolved by a cross-diffusion reaction–diffusion system that enforces inter-class competition and exclusivity. A differentiable implicit–explicit (IMEX) scheme treats stiff diffusion implicitly while handling nonlinear reaction and event terms explicitly, followed by projection back to the simplex. To further enhance global plausibility, we introduce a topology regularizer that simultaneously enforces centerline preservation and penalizes region overlaps. The approach is validated on synthetic datasets (Voronoi, Vessel) and a clinical dataset (UCSF-ALPTDG brain glioma). On synthetic benchmarks, our method achieves state-of-the-art accuracy (e.g., Voronoi-DSC: 95.70 ± 0.30 95.70\pm 0.30 and Vessel-DSC: 71.14 ± 0.25 71.14\pm 0.25 ) while preserving topology, and also demonstrates superior performance on the clinical dataset (UCSF-DSC: 65.37 ± 0.35 65.37\pm 0.35 ). By integrating PDE dynamics, topology-aware regularization, and differentiable solvers, this work establishes a principled path toward anatomy-to-anatomy generation for digital twins that are visually realistic, anatomically exclusive, and topologically consistent. Code will be made available upon acceptance. 1 Introduction Modeling the temporal evolution of anatomical structures is a central challenge in computational oncology and medical image analysis ( Ren et al. 2022 ; Lachinov et al. 2023 ) . Clinical imaging protocols routinely capture patient scans before and after treatment, or during disease progression, yet predictive models that generate plausible future anatomy from a baseline scan remain limited. The main reasons being scarce training datasets and lack in the existing mechanisms to enforce temporal consistency and anatomical plausibility in biological growth dynamics. There is a need to develop computational surrogates for individual patients that simulate trajectories of disease growth and therapeutic response, thereby informing personalized treatment planning and clinical decision support ( Katsoulakis et al. 2024 ; Kuang et al. 2024 ) . A potential solution to predicting growth trajectories is through generative AI approaches. Most existing generative modeling techniques in medical imaging, such as conditional generative adversarial networks (GANs) and diffusion models ( Ho et al. 2020 ; Dhariwal & Nichol 2021 ; Armanious et al. 2020 ) , operate primarily generate images from gaussian noise without considering any treatment paradigms as input. While these methods produce visually realistic outputs, they often fail to guarantee structural plausibility. To ensure structure correctness, several methods have been proposed that use conditioning mechanisms to generate images ( Zhang et al. 2023 ; Zhao et al. 2023 ) . Using anatomies as control has been shown to improve medical image generation ( Bhattacharya et al. 2024 ; Bhattacharya et al. 2025a ) . Recent methods can generate anatomically accurate post-treatment images from pre-treatment scans ( Bhattacharya et al. 2025b ; Liu et al. 2025 ) when conditioned with patient demographics, genomic markers, etc. The major limitations of these methods are that they are static in nature and cannot handle heterogenous inputs i.e., pre-treatment, pre-operative, post-operative scans, varying treatment schemes, genomic mutations. Figure 1: Existing baselines model only tumor growth and do not consider adjacent anatomy changes. Our proposed method, Anatomy-DT, models both anatomy and tumor growth conditioned on different treatment paradigms. In contrast, mechanistic models of tumor growth and tissue dynamics, often formulated as reaction–diffusion partial differential equations (PDEs) ( Martens et al. 2022 ; Yin et al. 2019 ; Metzcar et al. 2024 ) , provide interpretable dynamics grounded in biology. These models capture important features such as infiltration along tissue interfaces or mass effect on surrounding tissue. However, they are typically restricted to single-class tumor representations, lack explicit coupling with surrounding anatomy, and cannot be easily coupled with deep learning frameworks for end-to-end training. Moreover, they rarely account for global topological invariants, such as the connectivity of ventricles or the tree structure of vessels, which are critical to ensure clinically valid anatomical trajectories. Recently proposed works ( ( Bhattacharya et al. 2025a ) and Gupta et al. 2024 ) take topological constraints into account to ensure topological consistency in the generated images ( Bhattacharya et al. 2025a ; Gupta et al. 2024 ) . With the emergence of digital twins, mechanistic models conditioned on diverse treatment modalities have re-emerged as an important paradigm. Clinically, however, these treatments exert influence not only on the tumor itself but also on adjacent anatomical structures. Capturing the coupled evolution of multiple anatomies is particularly challenging, as it requires maintaining both spatial coherence and topological consistency across interacting tissues. Topology-constrained mechanistic models provide a principled framework to address this challenge, ensuring preservation of anatomical topology while enabling clinically faithful simulations. In this work, we introduce a novel framework, Anatomy-DT, an anatomy digital twin that bridges the gap between mechanistic PDEs and data-driven generative models. Our approach represents anatomy as a multi-class probability field u(x,t)\in\Delta^{K-1} , where K denotes the number of tissue classes and \Delta^{K-1} is the probability simplex. The temporal evolution of this field is governed by a cross-diffusion reaction–diffusion system , which naturally enforces competition between classes and exclusivity of tissue regions. Concretely, the governing PDE is formulated as \partial_{t}u_{i}(x,t)\;=\;\nabla\cdot\Bigg(\sum_{j=1}^{K}D_{ij}(u)\,\nabla u_{j}(x,t)\Bigg)\;+\;R_{i}(u) where D_{ij}(u) encodes both self-diffusion ( i=j ) and cross-diffusion ( i\neq j ) coefficients, and R_{i}(u) represents local reaction or growth terms (e.g., proliferation, atrophy, or event-driven changes). Such systems generalize classical reaction–diffusion equations Murray 2001 ; Turing 1990 ; Yankeelov et al. 2015 ; Jarrett et al. 2018 and have been studied extensively in multi-species dynamics Shigesada et al. 1979 ; Chen et al. 2021 ; Burger et al. 2020 . In the anatomical setting, this structure enforces sharp, mutually exclusive tissue boundaries without resorting to ad-hoc normalization. To achieve stable numerical integration within deep learning pipelines, we develop a differentiable implicit–explicit (IMEX) scheme Ascher et al. 1995 ; Hundsdorfer & Verwer 2013 that treats the stiff diffusion operator implicitly, while handling the nonlinear reaction and event terms explicitly. After each update, the solution is projected back onto the simplex, ensuring tissue probability conservation and anatomical plausibility. This design yields a numerically stable PDE layer that is fully compatible with backpropagation and scalable to medical images. To further guarantee anatomical plausibility beyond local PDE dynamics, we incorporate global topological priors. Anatomical structures are inherently complex, and their organization becomes further disrupted during tumor growth owing to processes such as angiogenesis, stromal remodeling, extracellular matrix degradation, edema formation, and mass effect, which collectively alter both local microarchitecture and global tissue geometry. In this work, we focus on modeling how anatomical structures deform in response to tumor growth. We propose a formulation that enforces topological correctness by combining centerline consistency for complex