MPGR-Net: multi-scale patch graph reasoning for weakly supervised pathological image segmentation Authors: Xu Zhang, Huaiju Ge, Yanling Wang, Chongqi Wei, Yang Liu, Xinbin Zhang, Cheng Li Published: 2026 ABSTRACT Weakly supervised pathological image segmentation aims to localize disease regions using coarse annotations (e.g., image-level labels), which significantly reduces annotation cost but introduces challenges such as imprecise localization and noisy predictions. In this paper, we propose multi-scale patch graph reasoning network (MPGR-Net), a novel framework that integrates multi-scale graph reasoning into transformer-based architectures to address these limitations. Specifically, we first employ a Vision Transformer encoder to extract global-aware patch representations. To explicitly model structured spatial dependencies that are often overlooked by standard self-attention, we introduce a multi-scale patch graph reasoning module, where patch tokens are treated as graph nodes and connected via predefined spatial adjacency at multiple scales. This design enables effective aggregation of local and long-range contextual information while maintaining computational efficiency. Furthermore, we incorporate a lightweight cross-attention mechanism to enhance global-to-local feature interaction through channel-wise modulation. Extensive experiments on pathological image segmentation benchmarks demonstrate that MPGR-Net consistently outperforms existing weakly supervised methods, achieving superior segmentation accuracy and robustness. Ablation studies further validate the effectiveness of the proposed graph reasoning module. Introduction Pathological image segmentation plays a crucial role in computer-aided diagnosis and clinical decision-making, as it enables precise localization of disease regions in high-resolution histopathology images [1, 2]. However, obtaining pixel-level annotations for pathological images is extremely labor-intensive, time-consuming, and requires expert knowledge. To alleviate this burden, weakly supervised segmentation, which relies on coarse annotations such as image-level labels, has attracted increasing attention in recent years. Despite its practical advantages, weakly supervised pathological image segmentation remains a challenging problem [3, 4]. The absence of fine-grained supervision often leads to incomplete object localization, noisy activation maps, and poor boundary delineation. Existing approaches typically rely on class activation maps (CAMs) or transformer-based attention mechanisms to infer spatial regions of interest [5–7]. However, these methods suffer from two major limitations. First, although Vision Transformers (ViTs) are effective in modeling long-range dependencies, their self-attention mechanism captures spatial relationships in an implicit manner, lacking explicit structural constraints. Second, most methods treat patch tokens independently or rely solely on attention weights, which may fail to enforce spatial consistency across neighboring regions. To address these limitations, we propose MPGR-Net, a novel framework that integrates multi-scale patch graph reasoning into transformer-based segmentation for weakly supervised pathological images. The key idea is to explicitly model spatial dependencies among patch tokens using structured graph representations, thereby enhancing both local coherence and global contextual understanding. Specifically, we first employ a Vision Transformer encoder to extract patch-level representations with global receptive fields. To overcome the limitations of implicit spatial modeling in self-attention, we construct a set of graphs over patch tokens, where each node corresponds to a patch and edges are defined based on spatial proximity at multiple scales. We then introduce a multi-scale graph reasoning module, which aggregates neighborhood information across different connectivity levels through graph neural network. This design enables the model to capture both fine-grained local structures and long-range spatial interactions in a unified framework. In addition, we incorporate a lightweight cross-attention-based global-to-local interaction module, where the global classification token is used to modulate patch features via channel-wise gating. This mechanism enhances the alignment between global semantics and local representations, leading to more accurate localization of pathological regions. Extensive experiments on pathological image segmentation benchmarks demonstrate that the proposed MPGR-Net consistently outperforms existing weakly supervised methods. In particular, our approach achieves superior performance in terms of segmentation accuracy, robustness, and localization quality. In summary, the main contributions of this work are as follows:We propose MPGR-Net, a novel transformer-based framework that introduces multi-scale patch graph reasoning for weakly supervised pathological image segmentation.We design a structured graph modeling approach that explicitly captures spatial dependencies among patch tokens across multiple scales, improving both local consistency and global context modeling. Our work is open-sourced on Github1. We propose MPGR-Net, a novel transformer-based framework that introduces multi-scale patch graph reasoning for weakly supervised pathological image segmentation. We propose MPGR-Net, a novel transformer-based framework that introduces multi-scale patch graph reasoning for weakly supervised pathological image segmentation. We design a structured graph modeling approach that explicitly captures spatial dependencies among patch tokens across multiple scales, improving both local consistency and global context modeling. Our work is open-sourced on Github1. We design a structured graph modeling approach that explicitly captures spatial dependencies among patch tokens across multiple scales, improving both local consistency and global context modeling. Our work is open-sourced on Github1. Related work Weakly supervised pathological image segmentation Weakly supervised pathological image segmentation has attracted increasing attention due to the high cost of dense pixel-level annotations in histopathology images. Early studies primarily rely on image-level labels to train segmentation models under weak supervision. A representative paradigm is Class Activation Mapping (CAM), which localizes discriminative regions by leveraging classification networks. However, CAM-based methods often suffer from incomplete activation, tending to highlight only the most discriminative parts of lesions while ignoring their full spatial extent [6, 8]. To address this limitation, numerous methods have been proposed to refine CAMs using additional constraints or learning strategies. Multiple Instance Learning (MIL) is widely adopted in weakly supervised medical image segmentation, where each image is treated as a bag of instances, and predictions are aggregated at the bag level [9, 10]. Although MIL improves robustness, it fails to explicitly model spatial dependencies among instances, leading to suboptimal segmentation boundaries. To further improve localization quality, pseudo-label refinement methods have been introduced. These approaches iteratively refine coarse predictions using self-training or region growing strategies [2, 11]. In parallel, consistency learning and self-training frameworks have been explored to enhance robustness under weak supervision [12]. Recently, transformer-based architectures have been introduced into weakly supervised segmentation due to their strong global modeling capability. Methods such as TransCAM leverage transformer attention to refine CAM outputs, improving spatial coverage [5]. Hybrid CNN-Transformer architectures further enhance feature representation by combining local convolutional features with global attention mechanisms [13, 14]. Despite these advances, most existing methods still rely on implicit attention mechanisms and lack explicit spatial structural modeling, which limits their performance in complex pathological structures. In contrast, our method introduces explicit multi-scale graph reasoning over patch tokens, enabling structured spatial modeling beyond attention-based approaches. Graph-based representation learning Graph-based representation learning provides a powerful framework for modeling non-Euclidean structured data. Graph Neural Networks (GNNs) extend deep learning to graph-structured inputs by performing message passing between connected nodes [15, 16]. In vision tasks, images can be represented as graphs where nodes correspond to pixels, superpixels, or patch embeddings, and edges encode spatial or semantic relationships. Early works apply graph learning to image segmentation by constructing superpixel graphs and performing graph neural network to refine region consistency [17, 18]. These methods demonstrate the effectiveness of explicit structure modeling but are limited by fixed graph construction and shallow architectures. With the rise of deep learning, graph-based methods have been integrated with CNN features to improve visual reasoning. For instance, Graph Convolutional Networks (GCNs) have been used for scene understanding and semantic segmentation by modeling contextual dependencies among regions [17, 19]. However, these approaches often rely on static or manually designed graph structures, limiting their adaptability. Recent studies have further explored hierarchical graph reasoning and local-to-global topology modeling for robust representation learning. For example, HIGDA [20] constructs hierarchical graph structures to capture local-to-global topological relationships for semi-supervised domain adaptation, while prototype-driven graph convolution methods enhance feature alignment through graph-based prototype propagation [21]. These works demonstrate the effectiveness of structured graph reasoning in modeling complex semantic dependencies, which further motivates our multi-scale patch graph reasoning framework. In medical imaging, graph-based methods have been explored to model anatomical relationships and tissue interactions [22, 23]. Despite their effectiveness, most existing approaches operate on a single-scale graph and fail to capture hierarchical spatial dependencies. More recently, efforts have been made to combine transformers with graph neural networks. Some studies incorporate attention mechanisms into MIL frameworks to model instance relationships [24]. Others attempt to unify Transformer or CNN backbones with graph reasoning modules for improved representation learning [16]. However, these methods still lack explicit multi-scale graph construction and efficient integration with patch-based transformer tokens. To address these limitations, we propose a multi-scale patch graph reasoning framework that explicitly constructs graphs at different spatial scales, enabling comprehensive modeling of both local and global dependencies in pathological images. The proposed approach Overview In this work, we propose a novel Multi-scale Patch Graph Reasoning Network (MPGR-Net) for image segmentation, which integrates transformer-based representation learning with structured graph reasoning as shown in Fig. 1. The core motivation is to explicitly model spatial dependencies among patch tokens while preserving the global contextual modeling capability of Vision Transformers.Fig. 1The architecture of MPGR-Net The architecture of MPGR-Net Given an input image \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{I} \in \mathbb{R}^{H \times W \times 3}$$\end{document}, the overall pipeline of the proposed framework can be formulated as: 1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{I} \xrightarrow{\mathcal{E}} \textbf{X} \xrightarrow{\mathcal{G}} \textbf{X}^{*} \xrightarrow{\mathcal{D}} \textbf{M},$$\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}$$\mathcal{E}$$\end{document} denotes the transformer encoder, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal{G}$$\end{document} represents the proposed multi-scale graph reasoning module, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal{D}$$\end{document} is the segmentation decoder. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{M}$$\end{document} denotes the final dense prediction. Token Representation. The encoder \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal{E}$$\end{document} first divides the image into non-overlapping patches and embeds them into a sequence of tokens:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{X} = \{x_{\mathrm{cls}}, x_1, x_2, \dots, x_N\} \in \mathbb{R}^{(N+1) \times C},$$\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}$$x_{\mathrm{cls}} \in \mathbb{R}^{C}$$\end{document} is the global classification token and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\{x_i\}_{i=1}^{N}$$\end{document} are patch tokens. Here, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$N = H_p \times W_p$$\end{document} denotes the number of patches. We separate the token sequence into:3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{X}_p = \{x_1, x_2, \dots, x_N\} \in \mathbb{R}^{N \times C},$$\end{document} which will be further processed for spatial reasoning. Global-to-Local Interaction. To incorporate global semantic guidance into local representations, we optionally introduce a cross-attention based interaction module:4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{X}_p\prime = \mathcal{A}(x_{\mathrm{cls}}, \textbf{X}_p),$$\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}$$\mathcal{A}(\cdot)$$\end{document} denotes the cross-attention operator. This module enhances patch tokens via channel-wise gating conditioned on the global token. Multi-scale Graph Reasoning. To explicitly model spatial relationships, we construct a set of graphs over patch tokens:5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal{G} = \{\textbf{A}^{(s)}\}_{s=1}^{S},$$\end{document} where each \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{A}^{(s)} \in \mathbb{R}^{N \times N}$$\end{document} represents a graph adjacency matrix at scale s. The patch tokens are then refined via multi-layer graph reasoning:6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{X}^{*} = \mathcal{G}(\textbf{X}_{p}^{\prime}),$$\end{document} which aggregates neighborhood information across multiple spatial scales. Decoder. The refined tokens \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{X}^{*}$$\end{document} are fed into a decoder \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal{D}$$\end{document} to generate patch-level predictions:7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{M}_p = \mathcal{D}(\textbf{X}^{*}),$$\end{document} which are further upsampled to pixel-level segmentation:8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{M} = \mathrm{Upsample}(\textbf{M}_p).$$\end{document} Training Objective. To enhance global consistency and suppress noisy activations, we employ a hybrid objective combining normalized global weighted pooling and focal regularization:9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal{L} = \mathrm{BCE}(\sigma(\textbf{y}), \textbf{y}_{\mathrm{gt}}),$$\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}$$\textbf{y}$$\end{document} is derived from the NGWP-Focal mechanism. Key Characteristics. The proposed framework has the following desirable properties:Structured Spatial Modeling: Graph reasoning explicitly captures spatial dependencies beyond standard self-attention.Multi-scale Context Aggregation: Multiple adjacency matrices enable flexible receptive fields.Global-Local Interaction: Cross-attention injects global semantics into local features. Structured Spatial Modeling: Graph reasoning explicitly captures spatial dependencies beyond standard self-attention. Structured Spatial Modeling: Graph reasoning explicitly captures spatial dependencies beyond standard self-attention. Multi-scale Context Aggregation: Multiple adjacency matrices enable flexible receptive fields. Multi-scale Context Aggregation: Multiple adjacency matrices enable flexible receptive fields. Global-Local Interaction: Cross-attention injects global semantics into local features. Global-Local Interaction: Cross-attention injects global semantics into local features. Global-to-local interaction Although Vision Transformers are capable of modeling long-range dependencies via self-attention, the interaction between the global semantic token and local patch tokens remains implicit. To explicitly inject global contextual information into local representations, we introduce a Global-to-Local Interaction (GTLI) module. The classification token \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_{\mathrm{cls}}$$\end{document} encodes high-level semantic information of the entire image, while patch tokens \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{X}_p$$\end{document} primarily capture localized patterns. However, standard transformer layers do not explicitly enforce global-to-local modulation. This may limit the model’s ability to refine ambiguous or weakly activated regions. To address this issue, we design a lightweight cross-attention mechanism that enables the global token to selectively enhance patch features through channel-wise gating. Given patch tokens \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{X}_p \in \mathbb{R}^{N \times C}$$\end{document} and the global token \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_{\mathrm{cls}} \in \mathbb{R}^{C}$$\end{document}, we first project them into query, key, and value spaces:10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{Q} = W_q x_{\mathrm{cls}} \in \mathbb{R}^{C},$$\end{document}11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ \textbf{K}, \textbf{V} = \mathrm{split}(W_{kv} \textbf{X}_p), \quad \textbf{K}, \textbf{V} \in \mathbb{R}^{N \times C},$$\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}$$W_q \in \mathbb{R}^{C \times C}$$\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}$$W_{kv} \in \mathbb{R}^{C \times 2C}$$\end{document} are learnable projection matrices. We then perform multi-head cross-attention:12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{r} = \mathrm{Attention}(\textbf{Q}, \textbf{K}, \textbf{V}),$$\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}$$\textbf{r} \in \mathbb{R}^{C}$$\end{document} represents a global response vector summarizing patch-level information conditioned on the CLS token. Channel-wise Gating. To transform the global response into modulation signals, we introduce a gating function:13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{g} = \sigma(W_r \textbf{r}),$$\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}$$W_r \in \mathbb{R}^{C \times C}$$\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}$$\sigma(\cdot)$$\end{document} denotes the sigmoid activation function. The gate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{g} \in [0,1]^C$$\end{document} acts as a channel-wise importance weight. Feature Modulation. The refined patch tokens are computed via residual gating:14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{X}_{p}^{\prime} = \textbf{X}_p + \textbf{g} \odot \textbf{X}_p,$$\end{document} where ⊙ denotes element-wise multiplication with broadcasting along the spatial dimension. Finally, a Layer Normalization is applied:15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{X}_{p}^{\prime} = \mathrm{LayerNorm}(\textbf{X}_{p}^{\prime}).$$\end{document} Multi-scale patch graph reasoning While transformer-based self-attention captures global dependencies, it lacks explicit modeling of structured spatial relationships. In dense prediction tasks such as medical image segmentation, spatial continuity and neighborhood consistency are crucial. To address this limitation, we propose a Multi-scale Patch Graph Reasoning (MPGR) module that explicitly models interactions among patch tokens via graph neural network over multiple spatial scales. Graph Construction. Given patch tokens \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{X}_p \in \mathbb{R}^{N \times C}$$\end{document} arranged on a 2D grid of size \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H_p \times W_p$$\end{document}, we construct a set of graphs:16\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal{G} = \{\textbf{A}^{(s)} \mid s = 1, 2, \dots, S \},$$\end{document} where each \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{A}^{(s)} \in \mathbb{R}^{N \times N}$$\end{document} denotes the adjacency matrix at scale s. Each adjacency matrix is defined based on spatial connectivity:17\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$A^{(s)}_{ij} =\begin{cases}1, & \text{if node } j \in \mathcal{N}^{(s)}(i) \\0, & \mathrm{otherwise}\end{cases}$$\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}$$\mathcal{N}^{(s)}(i)$$\end{document} represents the neighborhood of node i under connectivity level s (e.g., 4-, 8-, or extended neighborhoods). To ensure numerical stability and balanced aggregation, we normalize each adjacency matrix as:18\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{\textbf{A}}^{(s)} = \textbf{D}^{-1} \textbf{A}^{(s)},$$\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}$$\textbf{D}$$\end{document} is the degree matrix. Multi-scale Graph Neural Network. Given the set of normalized adjacency matrices, we perform graph reasoning by aggregating features from multiple spatial scales. For the l-th layer, the update rule is defined as:19\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{X}^{(l+1)} = \textbf{X}^{(l)} + W_{\mathrm{self}} \textbf{X}^{(l)} + \sum_{s=1}^{S} W_s \left( \hat{\textbf{A}}^{(s)} \textbf{X}^{(l)} \right),$$\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}$$W_{\mathrm{self}} \in \mathbb{R}^{C \times C}$$\end{document} models self-feature transformation,\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$W_s \in \mathbb{R}^{C \times C}$$\end{document} is the learnable projection for scale s. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$W_{\mathrm{self}} \in \mathbb{R}^{C \times C}$$\end{document} models self-feature transformation, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$W_{\mathrm{self}} \in \mathbb{R}^{C \times C}$$\end{document} models self-feature transformation, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$W_s \in \mathbb{R}^{C \times C}$$\end{document} is the learnable projection for scale s. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$W_s \in \mathbb{R}^{C \times C}$$\end{document} is the learnable projection for scale s. Layer Stacking. To progressively refine representations, we stack multiple graph reasoning layers: 20\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{X}^{(L)} = \mathcal{F}_{\mathrm{GNN}}(\textbf{X}_{p}^{\prime}),$$\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}$$\mathcal{F}_{\mathrm{GNN}}(\cdot)$$\end{document} denotes the stacked multi-scale graph neural network blocks. Each layer is equipped with residual connections, Layer Normalization, and dropout:21\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{X}^{(l+1)} = \textbf{X}^{(l)} + \mathrm{Dropout}(\mathcal{H}(\mathrm{LayerNorm}(\textbf{X}^{(l)}))),$$\end{document} which improves training stability and prevents overfitting. Efficient Adjacency Construction. Unlike dynamic graph learning methods, we construct adjacency matrices based on fixed spatial layouts, which brings two advantages:Deterministic structure: No additional parameters are required.Computational efficiency: Adjacency matrices can be precomputed and cached. Deterministic structure: No additional parameters are required. Deterministic structure: No additional parameters are required. Computational efficiency: Adjacency matrices can be precomputed and cached. Computational efficiency: Adjacency matrices can be precomputed and cached. Furthermore, different connectivity levels (e.g., 4, 8, 12, 16, 24) provide flexible control over receptive fields without increasing model depth. Decoder After multi-scale graph reasoning, the refined patch tokens encode rich contextual and structural information. To generate dense segmentation maps, we adopt a mask transformer decoder, which models the interactions between patch tokens and class embeddings to produce segmentation masks in a learnable and global manner. Input Representation. The graph reasoning module outputs refined patch features:22\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{X}^{*} \in \mathbb{R}^{N \times C},$$\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}$$N = H_p \times W_p$$\end{document} denotes the number of patches. In addition, we introduce a set of learnable class embeddings:23\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{E} \in \mathbb{R}^{K \times C},$$\end{document} where K is the number of semantic classes. Token-Class Interaction. The mask transformer models the interaction between patch tokens and class embeddings via attention:24\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{Z} = \mathrm{Attention}(\textbf{E}, \textbf{X}^{*}),$$\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}$$\textbf{Z} \in \mathbb{R}^{K \times C}$$\end{document} encodes class-specific responses aggregated from patch tokens. Alternatively, the interaction can be interpreted as computing the similarity between patch tokens and class embeddings:25\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{M}_p = \textbf{E} \cdot (\textbf{X}^{*})^{\top},$$\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}$$\textbf{M}_p \in \mathbb{R}^{K \times N}$$\end{document} represents patch-level mask logits. Patch-level Mask Prediction. The predicted masks are reshaped into spatial maps: 26\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{M}_p \in \mathbb{R}^{K \times H_p \times W_p}.$$\end{document} This formulation enables each class embedding to attend to relevant patch tokens, producing class-specific segmentation masks. Upsampling to Pixel Space. The patch-level masks are upsampled to the original image resolution via bilinear interpolation:27\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{M} = \mathrm{Interp}(\textbf{M}_p, H, W),$$\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}$$\textbf{M} \in \mathbb{R}^{K \times H \times W}$$\end{document} is the final segmentation output. Training objective To obtain image-level predictions under weak supervision, we apply the Normalized Global Weighted Pooling (NGWP) operator on the patch-level mask logits produced by the decoder. Given the coarse prediction map \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{M}_p \in \mathbb{R}^{K \times H_p \times W_p}$$\end{document}, we first reshape it into:28\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{M}_p \in \mathbb{R}^{K \times N}, \quad N = H_p \times W_p.$$\end{document} The corresponding probability map is computed via softmax along the class dimension:29\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{P} = \textrm{Softmax}(\textbf{M}_p),$$\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}$$\textbf{P} \in \mathbb{R}^{K \times N}$$\end{document}. The final image-level prediction is defined as:30\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{y} = \textbf{y}_{\textrm{ngwp}} + \beta \, \textbf{y}_{\textrm{focal}},$$\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}$$\textbf{y}_{\textrm{ngwp}}$$\end{document} denotes the NGWP aggregation and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{y}_{\textrm{focal}}$$\end{document} represents the focal regularization term. Normalized Global Weighted Pooling. The NGWP operator performs confidence-aware aggregation over spatial locations:31\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{y}_{\textrm{ngwp}} =\frac{\sum_{i=1}^{N} \textbf{M}_p^{(i)} \odot \textbf{P}^{(i)}}{1 + \sum_{i=1}^{N} \textbf{P}^{(i)}},$$\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}$$\textbf{M}_p^{(i)} \in \mathbb{R}^{K}$$\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}$$\textbf{P}^{(i)} \in \mathbb{R}^{K}$$\end{document} denote the logits and probabilities at spatial position i, and ⊙ denotes element-wise multiplication. Focal Regularization. To encourage sparse and discriminative activations, a focal-style regularization term is introduced:32\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{y}_{\textrm{focal}} =(1 - \bar{\textbf{P}})^3 \odot \log(\lambda + \bar{\textbf{P}}),$$\end{document} where 33\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\bar{\textbf{P}} = \frac{1}{N} \sum_{i=1}^{N} \textbf{P}^{(i)}$$\end{document} is the spatial mean probability per class, and λ is a small constant for numerical stability. Loss Function. We apply a binary cross-entropy loss on the sigmoid-activated outputs: 34\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal{L} = \mathrm{BCE}(\sigma(\textbf{y}), \textbf{y}_{\mathrm{gt}}),$$\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}$$\textbf{y}_{\mathrm{gt}} \in \mathbb{R}^{K}$$\end{document} denotes image-level labels. Experiments In this section, we conduct a comprehensive evaluation of the proposed MPGR-Net on two publicly available weakly supervised histopathology segmentation benchmarks. We first introduce the datasets and implementation details, followed by comparisons with state-of-the-art methods and quantitative evaluation metrics. Datasets We evaluate our method on two representative histopathology datasets covering lung and breast cancer tissue segmentation under weak supervision. LUAD-HistoSeg. This dataset consists of lung adenocarcinoma (LUAD) histopathology patches with image-level annotations [2]. It includes four tissue categories: Tumor Epithelial (TE), Tumor-Associated Stroma (TAS), Lymphocyte (LYM), and Necrosis (NEC). In total, 17,285 image patches of size 224 × 224 are provided, among which 16,678 are used for training with only image-level labels, while 300 validation and 307 test samples are annotated with pixel-level labels for evaluation. BCSS-WSSS. The BCSS-WSSS dataset is derived from H&E-stained breast cancer whole-slide images [2], containing four classes: Tumor (TUM), Stroma (STR), Lymphocytic infiltrate (LYM), and Necrosis (NEC). The dataset includes 31,826 patches of size 224 × 224, split into 23,422 training images, 3,418 validation images, and 4,986 test images, where pixel-level annotations are available for validation and testing. Implementation details All experiments are implemented in PyTorch and conducted on a single NVIDIA RTX 3090Ti GPU. The input images are resized to 224 × 224 and divided into patches of size P=16. We adopt a Vision Transformer backbone (ViT-S/16) with an embedding dimension of D=384 and 12 transformer layers. The proposed multi-scale graph reasoning module is stacked with L=8 layers and uses multiple spatial connectivities (e.g., \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\{4, 8, 12, 16\}$$\end{document}). A mask transformer decoder is employed to generate patch-level segmentation masks. The model is optimized using the SGD optimizer with an initial learning rate of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1 \times 10^{-2}$$\end{document}. The batch size is set to 32, and the model is trained for 10 epochs. Comparison methods We compare MPGR-Net with several representative weakly supervised histopathology segmentation methods:HistoSegNet [25]: A CNN-based framework for weakly supervised tissue segmentation that leverages patch-level annotations to localize diagnostically relevant regions.OEEM [26]: A weakly supervised segmentation method that emphasizes reliable regions via online example mining to mitigate noisy supervision.WSSS-Tissue [2]: A two-stage framework that generates pseudo masks from image-level labels and refines them for segmentation.TPRO [27]: A text-prompt-based approach that incorporates language priors to enhance pseudo-label quality.PistoSeg [28]: A synthesis-based method that converts weak supervision into a fully supervised paradigm via mosaic augmentation and consistency learning.HisynSeg [29]: A hybrid supervision framework that combines image synthesis and self-supervised consistency to improve segmentation performance.PLIMSeg [30]: A CLIP-based method that leverages pathology-specific language–image alignment to generate more complete activation maps. HistoSegNet [25]: A CNN-based framework for weakly supervised tissue segmentation that leverages patch-level annotations to localize diagnostically relevant regions. HistoSegNet [25]: A CNN-based framework for weakly supervised tissue segmentation that leverages patch-level annotations to localize diagnostically relevant regions. OEEM [26]: A weakly supervised segmentation method that emphasizes reliable regions via online example mining to mitigate noisy supervision. OEEM [26]: A weakly supervised segmentation method that emphasizes reliable regions via online example mining to mitigate noisy supervision. WSSS-Tissue [2]: A two-stage framework that generates pseudo masks from image-level labels and refines them for segmentation. WSSS-Tissue [2]: A two-stage framework that generates pseudo masks from image-level labels and refines them for segmentation. TPRO [27]: A text-prompt-based approach that incorporates language priors to enhance pseudo-label quality. TPRO [27]: A text-prompt-based approach that incorporates language priors to enhance pseudo-label quality. PistoSeg [28]: A synthesis-based method that converts weak supervision into a fully supervised paradigm via mosaic augmentation and consistency learning. PistoSeg [28]: A synthesis-based method that converts weak supervision into a fully supervised paradigm via mosaic augmentation and consistency learning. HisynSeg [29]: A hybrid supervision framework that combines image synthesis and self-supervised consistency to improve segmentation performance. HisynSeg [29]: A hybrid supervision framework that combines image synthesis and self-supervised consistency to improve segmentation performance. PLIMSeg [30]: A CLIP-based method that leverages pathology-specific language–image alignment to generate more complete activation maps. PLIMSeg [30]: A CLIP-based method that leverages pathology-specific language–image alignment to generate more complete activation maps. Evaluation metrics We adopt the widely used Intersection over Union (IoU) and mean IoU (mIoU) to evaluate segmentation performance. For each class c, the IoU is defined as:35\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textrm{IoU}_c = \frac{|\hat{M}_c \cap M_c|}{|\hat{M}_c \cup M_c|},$$\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}$$\hat{M}_c$$\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}$$M_c$$\end{document} denote the predicted and ground-truth pixel sets, respectively. The mean IoU is computed as:36\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textrm{mIoU} = \frac{1}{C} \sum_{c=1}^{C} \textrm{IoU}_c,$$\end{document} where C is the number of classes. Experiment results and analysis Comparison between MPGR-Net and Previous Works Table 1 reports the quantitative comparison between the proposed MPGR-Net and several state-of-the-art weakly supervised histopathology segmentation methods on two benchmarks, i.e., LUAD-HistoSeg and BCSS-WSSS.Table 1Comparison between MPGR-Net and previous works on LUAD-HistoSeg and BCSS-WSSSMethodLUAD-HistoSegBCSS-WSSSTENECLYMTASmIoUTUMSTRLYMNECmIoUHistoSegNet45.5936.3058.2850.8147.7533.1446.4629.051.9127.64WSSS-Tissue74.5464.4852.5458.6762.5573.8168.2855.3858.3464.08OEEM73.8170.4971.8969.4871.4274.8664.6848.9161.0362.37TPRO74.8078.2071.1069.8073.5076.2065.1054.2062.8064.60PistoSeg77.1578.5070.8370.0274.1281.1673.9856.2666.7369.53HisynSeg78.2378.4677.1571.1976.2581.6474.6760.1566.4870.73PLIMSeg78.7281.1073.5472.1476.3881.3975.4160.3367.1671.07MPGR-Net78.5079.6575.3772.2276.4380.1974.7863.0569.7071.95 Table 1Comparison between MPGR-Net and previous works on LUAD-HistoSeg and BCSS-WSSSMethodLUAD-HistoSegBCSS-WSSSTENECLYMTASmIoUTUMSTRLYMNECmIoUHistoSegNet45.5936.3058.2850.8147.7533.1446.4629.051.9127.64WSSS-Tissue74.5464.4852.5458.6762.5573.8168.2855.3858.3464.08OEEM73.8170.4971.8969.4871.4274.8664.6848.9161.0362.37TPRO74.8078.2071.1069.8073.5076.2065.1054.2062.8064.60PistoSeg77.1578.5070.8370.0274.1281.1673.9856.2666.7369.53HisynSeg78.2378.4677.1571.1976.2581.6474.6760.1566.4870.73PLIMSeg78.7281.1073.5472.1476.3881.3975.4160.3367.1671.07MPGR-Net78.5079.6575.3772.2276.4380.1974.7863.0569.7071.95 Comparison between MPGR-Net and previous works on LUAD-HistoSeg and BCSS-WSSS On the LUAD-HistoSeg dataset, MPGR-Net achieves an overall mIoU of 76.43%, which is competitive with the best-performing methods. Specifically, MPGR-Net demonstrates strong performance across multiple tissue categories, achieving 78.50% on TE, 79.65% on NEC, 75.37% on LYM, and 72.22% on TAS. Compared with earlier approaches such as HistoSegNet and WSSS-Tissue, MPGR-Net shows substantial improvements, indicating the effectiveness of multi-scale graph reasoning in capturing complex tissue structures. Although PLIMSeg achieves a slightly lower or comparable mIoU (76.38%), MPGR-Net exhibits more balanced performance across categories, particularly improving over PLIMSeg on LYM (+1.83%) and TAS (+0.08%), suggesting better generalization across heterogeneous tissue patterns. On the more challenging BCSS-WSSS dataset, MPGR-Net achieves the best overall performance with an mIoU of 71.95%, outperforming all competing methods. Notably, MPGR-Net significantly improves the segmentation of difficult categories such as LYM and NEC, achieving 63.05% and 69.70%, respectively. Compared with the previous best method PLIMSeg, MPGR-Net yields a notable gain of +0.88% in mIoU, while also improving LYM by +2.72% and NEC by +2.54%. These results highlight the superiority of MPGR-Net in modeling complex inter-class relationships and alleviating the ambiguity caused by weak supervision. Overall, the superior performance of MPGR-Net can be attributed to its ability to model long-range dependencies and multi-scale contextual relationships via graph-based reasoning. By integrating hierarchical patch interactions, MPGR-Net alleviates the limitations of conventional CAM-based methods, such as fragmented activation and inaccurate boundaries. The consistent improvements across two datasets validate the robustness and generalization capability of the proposed framework. Ablation Study To validate the effectiveness of each component in MPGR-Net, we conduct ablation experiments on two benchmarks, i.e., LUAD-HistoSeg and BCSS-WSSS. Specifically, we evaluate two degraded variants: (1) without GTLI, which removes the Global-to-Local Interaction module and retains only the Multi-scale Patch Graph Reasoning (MPGR), and (2) without MPGR, which removes the graph reasoning module and keeps only the Global-to-Local Interaction (GTLI). As shown in Table 2, the full MPGR-Net achieves the best overall performance on both datasets, with mIoU scores of 76.43% on LUAD-HistoSeg and 71.65% on BCSS-WSSS. Removing either component consistently degrades performance, demonstrating that both MPGR and GTLI are indispensable for accurate tissue segmentation.Table 2Ablation study on LUAD-HistoSeg and BCSS-WSSSMethodLUAD-HistoSegBCSS-WSSSTENECLYMTASmIoUTUMSTRLYMNECmIoUMPGR-Net78.5079.6575.3772.2276.4380.2674.2661.8170.2971.65without GTLI78.6174.8873.5373.4075.1180.3972.7061.2669.1770.88without MPGR75.9577.6472.8471.8774.5780.2273.1658.5668.1970.03 Table 2Ablation study on LUAD-HistoSeg and BCSS-WSSSMethodLUAD-HistoSegBCSS-WSSSTENECLYMTASmIoUTUMSTRLYMNECmIoUMPGR-Net78.5079.6575.3772.2276.4380.2674.2661.8170.2971.65without GTLI78.6174.8873.5373.4075.1180.3972.7061.2669.1770.88without MPGR75.9577.6472.8471.8774.5780.2273.1658.5668.1970.03 Ablation study on LUAD-HistoSeg and BCSS-WSSS When removing GTLI (i.e., only using MPGR), the performance drops from 76.43% to 75.83% on LUAD-HistoSeg and more significantly from 71.65% to 69.28% on BCSS-WSSS. Notably, substantial degradation is observed in challenging categories such as LYM (from 61.81% to 57.00%) and NEC (from 70.29% to 66.80%) on BCSS-WSSS. This indicates that GTLI plays a crucial role in bridging global semantic context and local patch features, which is particularly important under weak supervision where discriminative cues are incomplete. Interestingly, the variant without GTLI achieves slightly higher scores on certain classes (e.g., NEC on LUAD-HistoSeg), suggesting that pure graph reasoning can still capture strong local structures. However, the overall decline in mIoU demonstrates that the lack of global guidance leads to suboptimal feature alignment and reduced robustness. When removing MPGR (i.e., only using GTLI), the performance further decreases to 75.38% on LUAD-HistoSeg and 68.71% on BCSS-WSSS. Compared to the full model, this variant shows consistent degradation across almost all categories. In particular, categories requiring fine-grained structural modeling, such as LYM and NEC, suffer noticeable drops (e.g., NEC decreases from 70.29% to 65.07% on BCSS-WSSS). These results demonstrate that MPGR is essential for modeling spatial dependencies and multi-scale contextual relationships among patches. By comparing the two variants, it is evident that removing MPGR leads to a larger performance drop than removing GTLI, indicating that graph-based reasoning contributes more to capturing structural information. However, the best performance is only achieved when both modules are jointly integrated. MPGR focuses on modeling local-to-global spatial relationships via graph interactions, while GTLI enhances feature representations through global semantic guidance. Their combination enables more complete and coherent segmentation, especially in complex histopathological images. Overall, the ablation results confirm that both MPGR and GTLI play complementary roles in the proposed framework. MPGR improves spatial consistency and boundary completeness through multi-scale reasoning, while GTLI alleviates the limitations of weak supervision by injecting global contextual priors. The synergy between these two components leads to consistent improvements across datasets and categories, validating the design of MPGR-Net. Effect of Multi-scale Connectivity on Graph Reasoning Table 3 presents the performance of MPGR-Net under different multi-scale connectivity configurations. We observe that the choice of connectivity levels has a significant impact on segmentation performance across both datasets.Table 3Effect of different multi-scale connectivity configurations on LUAD-HistoSeg and BCSS-WSSSMPGR-NetLUAD-HistoSegBCSS-WSSSTENECLYMTASmIoUTUMSTRLYMNECmIoU(4, 8, 12)77.3177.6975.4371.8575.5779.5373.7059.9369.5870.69(8, 12, 16)78.1478.3274.9672.6276.0180.2073.9159.5568.1870.46(4, 8, 12, 16)78.5079.6575.3772.2276.4379.4873.4759.9568.6870.39(8, 12, 16, 20)77.1978.7675.3672.9276.0680.1574.3261.1369.1171.18(4, 8, 12, 16, 20)74.9979.4873.2269.9874.4279.7973.4560.2168.4270.47(8, 12, 16, 20, 24)77.5677.6874.1672.5875.4979.6273.1963.2967.5870.92(4, 8, 12, 16, 20, 24)71.4375.0173.0266.2271.4280.2674.2661.8170.2971.65(8, 12, 16, 20, 24, 28)76.9680.3574.4572.5676.0879.8373.9761.6968.2070.92(4, 8, 12, 16, 20, 24, 28)76.7877.6173.8771.0774.8379.0472.8362.5869.4870.98 Table 3Effect of different multi-scale connectivity configurations on LUAD-HistoSeg and BCSS-WSSSMPGR-NetLUAD-HistoSegBCSS-WSSSTENECLYMTASmIoUTUMSTRLYMNECmIoU(4, 8, 12)77.3177.6975.4371.8575.5779.5373.7059.9369.5870.69(8, 12, 16)78.1478.3274.9672.6276.0180.2073.9159.5568.1870.46(4, 8, 12, 16)78.5079.6575.3772.2276.4379.4873.4759.9568.6870.39(8, 12, 16, 20)77.1978.7675.3672.9276.0680.1574.3261.1369.1171.18(4, 8, 12, 16, 20)74.9979.4873.2269.9874.4279.7973.4560.2168.4270.47(8, 12, 16, 20, 24)77.5677.6874.1672.5875.4979.6273.1963.2967.5870.92(4, 8, 12, 16, 20, 24)71.4375.0173.0266.2271.4280.2674.2661.8170.2971.65(8, 12, 16, 20, 24, 28)76.9680.3574.4572.5676.0879.8373.9761.6968.2070.92(4, 8, 12, 16, 20, 24, 28)76.7877.6173.8771.0774.8379.0472.8362.5869.4870.98 Effect of different multi-scale connectivity configurations on LUAD-HistoSeg and BCSS-WSSS On LUAD-HistoSeg, the configuration (4, 8, 12, 16) achieves the best overall performance with an mIoU of 76.43%, indicating that combining both local (4, 8) and mid-range (12, 16) neighborhood structures provides the most effective spatial modeling. In contrast, using either fewer scales (e.g., (4, 8, 12)) or overly large receptive fields (e.g., (4, 8, 12, 16, 20, 24)) leads to performance degradation. This suggests that an appropriate balance between local detail preservation and global context aggregation is crucial. On BCSS-WSSS, the best performance is achieved by (4, 8, 12, 16, 20, 24) with an mIoU of 71.65%. Compared with LUAD-HistoSeg, BCSS-WSSS benefits more from larger connectivity ranges, likely due to the more complex tissue structures and higher inter-class variability in breast cancer histopathology images. Effect of Connectivity Scale. From the results, several important observations can be made:Multi-scale modeling is essential. Configurations with multiple connectivity levels consistently outperform single-range or limited-scale settings, demonstrating the necessity of capturing spatial dependencies at different receptive fields.Moderate connectivity yields optimal performance. Excessively large connectivity (e.g., including 24 or 28 neighbors) may introduce noisy or less relevant long-range interactions, which can degrade performance, particularly on LUAD-HistoSeg.Dataset-dependent behavior. LUAD-HistoSeg favors relatively compact connectivity ranges, while BCSS-WSSS benefits from broader spatial context. This highlights the adaptability of the proposed multi-scale graph reasoning framework. Multi-scale modeling is essential. Configurations with multiple connectivity levels consistently outperform single-range or limited-scale settings, demonstrating the necessity of capturing spatial dependencies at different receptive fields. Multi-scale modeling is essential. Configurations with multiple connectivity levels consistently outperform single-range or limited-scale settings, demonstrating the necessity of capturing spatial dependencies at different receptive fields. Moderate connectivity yields optimal performance. Excessively large connectivity (e.g., including 24 or 28 neighbors) may introduce noisy or less relevant long-range interactions, which can degrade performance, particularly on LUAD-HistoSeg. Moderate connectivity yields optimal performance. Excessively large connectivity (e.g., including 24 or 28 neighbors) may introduce noisy or less relevant long-range interactions, which can degrade performance, particularly on LUAD-HistoSeg. Dataset-dependent behavior. LUAD-HistoSeg favors relatively compact connectivity ranges, while BCSS-WSSS benefits from broader spatial context. This highlights the adaptability of the proposed multi-scale graph reasoning framework. Dataset-dependent behavior. LUAD-HistoSeg favors relatively compact connectivity ranges, while BCSS-WSSS benefits from broader spatial context. This highlights the adaptability of the proposed multi-scale graph reasoning framework. Overall, these results validate the effectiveness of the proposed multi-scale graph construction strategy. By integrating multiple neighborhood structures, MPGR-Net is able to balance local consistency and global context modeling. Moreover, the results suggest that selecting appropriate connectivity scales is important for achieving optimal performance, depending on the characteristics of the dataset. Although the proposed MPGR module adopts predefined spatial adjacency for efficient and stable graph construction, irregular pathological structures may exhibit semantic correlations beyond spatial proximity. Multi-scale connectivity partially alleviates this issue by enabling broader contextual aggregation across different receptive fields. In future work, we plan to investigate adaptive graph construction strategies, such as feature similarity-based dynamic edge generation and semantic affinity learning, to better model complex tissue morphology. Effect of Graph Reasoning Depth Table 4 reports the performance of MPGR-Net with different graph reasoning depths l. We observe that increasing the number of graph layers initially improves performance, but excessive depth leads to performance degradation.Table 4Effect of graph reasoning depth l on LUAD-HistoSeg and BCSS-WSSSMPGR-NetLUAD-HistoSegBCSS-WSSSTENECLYMTASmIoUTUMSTRLYMNECmIoUl = 476.5779.4973.3371.9275.3379.5773.1659.7967.8770.10l = 676.7578.4672.6971.4774.8480.4473.9360.0568.2970.68l = 778.6479.2073.5973.0276.1179.2873.6361.4267.6170.48l = 878.5079.6575.3772.2276.4380.2674.2661.8170.2971.65l = 977.9278.8873.5972.5475.7380.9874.2760.9368.1871.09l = 1076.6478.2771.6671.6274.5179.4573.2162.6865.4570.20l = 1278.0078.8873.3272.2675.6179.9873.7461.8468.5871.03 Table 4Effect of graph reasoning depth l on LUAD-HistoSeg and BCSS-WSSSMPGR-NetLUAD-HistoSegBCSS-WSSSTENECLYMTASmIoUTUMSTRLYMNECmIoUl = 476.5779.4973.3371.9275.3379.5773.1659.7967.8770.10l = 676.7578.4672.6971.4774.8480.4473.9360.0568.2970.68l = 778.6479.2073.5973.0276.1179.2873.6361.4267.6170.48l = 878.5079.6575.3772.2276.4380.2674.2661.8170.2971.65l = 977.9278.8873.5972.5475.7380.9874.2760.9368.1871.09l = 1076.6478.2771.6671.6274.5179.4573.2162.6865.4570.20l = 1278.0078.8873.3272.2675.6179.9873.7461.8468.5871.03 Effect of graph reasoning depth l on LUAD-HistoSeg and BCSS-WSSS On LUAD-HistoSeg, the performance steadily increases from l=4 to l=8, achieving the best mIoU of 76.43% at l=8. Beyond this point, the performance starts to decline (e.g., 75.73% at l=9 and 74.51% at l=10), indicating that overly deep graph propagation may introduce redundant or noisy information. A similar trend is observed on BCSS-WSSS. The performance improves as the depth increases, reaching the highest mIoU of 71.65% at l=8. Although deeper configurations (e.g., l=9 or l=12) remain competitive, they do not provide further gains and may even slightly degrade performance. Effect of Graph Depth. The results reveal several important insights:Progressive improvement with increasing depth. Increasing the number of graph layers enhances feature propagation and enables better modeling of long-range spatial dependencies.Existence of an optimal depth. Both datasets achieve the best performance at l=8, suggesting that there exists a balance between sufficient contextual aggregation and over-smoothing.Over-smoothing in deep graph layers. When the depth becomes too large, node features tend to become overly similar due to repeated message passing, leading to degraded discriminability. Progressive improvement with increasing depth. Increasing the number of graph layers enhances feature propagation and enables better modeling of long-range spatial dependencies. Progressive improvement with increasing depth. Increasing the number of graph layers enhances feature propagation and enables better modeling of long-range spatial dependencies. Existence of an optimal depth. Both datasets achieve the best performance at l=8, suggesting that there exists a balance between sufficient contextual aggregation and over-smoothing. Existence of an optimal depth. Both datasets achieve the best performance at l=8, suggesting that there exists a balance between sufficient contextual aggregation and over-smoothing. Over-smoothing in deep graph layers. When the depth becomes too large, node features tend to become overly similar due to repeated message passing, leading to degraded discriminability. Over-smoothing in deep graph layers. When the depth becomes too large, node features tend to become overly similar due to repeated message passing, leading to degraded discriminability. Overall, these results demonstrate that an appropriate graph depth is critical for effective representation learning. A moderate number of layers (e.g., l=8) allows MPGR-Net to capture rich contextual information while preserving feature diversity. In subsequent experiments, we adopt l=8 as the default setting. Effect of the Number of Attention Heads Table 5 reports the impact of the number of attention heads h on segmentation performance. We observe that the choice of h plays an important role in balancing feature diversity and representation stability.Table 5Effect of the number of attention heads h on LUAD-HistoSeg and BCSS-WSSSMPGR-NetLUAD-HistoSegBCSS-WSSSTENECLYMTASmIoUTUMSTRLYMNECmIoUh = 177.7077.7673.9071.6275.2579.1573.9062.9468.8971.22h = 277.7179.3173.3172.0775.6080.1974.7863.0569.7071.95h = 378.579.6575.3772.2276.4380.2674.2661.8170.2971.65h = 478.0878.8873.6572.3275.7380.5273.8762.4867.2271.02h = 676.2180.3576.272.2976.2680.1774.0159.6768.6870.63 Table 5Effect of the number of attention heads h on LUAD-HistoSeg and BCSS-WSSSMPGR-NetLUAD-HistoSegBCSS-WSSSTENECLYMTASmIoUTUMSTRLYMNECmIoUh = 177.7077.7673.9071.6275.2579.1573.9062.9468.8971.22h = 277.7179.3173.3172.0775.6080.1974.7863.0569.7071.95h = 378.579.6575.3772.2276.4380.2674.2661.8170.2971.65h = 478.0878.8873.6572.3275.7380.5273.8762.4867.2271.02h = 676.2180.3576.272.2976.2680.1774.0159.6768.6870.63 Effect of the number of attention heads h on LUAD-HistoSeg and BCSS-WSSS On LUAD-HistoSeg, the performance improves as the number of heads increases from h=1 to h=3, reaching the best mIoU of 76.43%. Increasing the number of heads beyond this point leads to a slight performance drop (e.g., 75.73% at h=4), although a relatively competitive result is still achieved at h=6 (76.26%). This suggests that a moderate number of attention heads is sufficient to capture diverse feature interactions without introducing excessive redundancy. On BCSS-WSSS, the best performance is achieved at h=2 with an mIoU of 71.95%, while h=3 also yields competitive results (71.65%). However, further increasing h does not lead to consistent improvements and may even degrade performance, particularly when h=6. Effect of Multi-head Attention. From the results, several key observations can be drawn:Moderate number of heads is optimal. Both datasets achieve strong performance when h is in the range of 2–3, indicating that a limited number of attention heads is sufficient to model diverse feature interactions.Too few heads limit representation capacity. When h=1, the model relies on a single attention subspace, which restricts its ability to capture heterogeneous patterns across different tissue structures.Too many heads introduce redundancy. Increasing the number of heads excessively may lead to redundant or noisy attention patterns, reducing the effectiveness of feature aggregation. Moderate number of heads is optimal. Both datasets achieve strong performance when h is in the range of 2–3, indicating that a limited number of attention heads is sufficient to model diverse feature interactions. Moderate number of heads is optimal. Both datasets achieve strong performance when h is in the range of 2–3, indicating that a limited number of attention heads is sufficient to model diverse feature interactions. Too few heads limit representation capacity. When h=1, the model relies on a single attention subspace, which restricts its ability to capture heterogeneous patterns across different tissue structures. Too few heads limit representation capacity. When h=1, the model relies on a single attention subspace, which restricts its ability to capture heterogeneous patterns across different tissue structures. Too many heads introduce redundancy. Increasing the number of heads excessively may lead to redundant or noisy attention patterns, reducing the effectiveness of feature aggregation. Too many heads introduce redundancy. Increasing the number of heads excessively may lead to redundant or noisy attention patterns, reducing the effectiveness of feature aggregation. Overall, these results demonstrate that the number of attention heads should be carefully selected to balance expressiveness and efficiency. A moderate setting (e.g., h=3) provides a good trade-off between representation diversity and stability, and is adopted as the default configuration in our experiments. Visualization analysis To further evaluate the effectiveness of the proposed MPGR-Net, we provide qualitative visualization results on LUAD-HistoSeg and BCSS-WSSS datasets, as shown in Fig. 2 and Fig. 3. We compare the activation maps generated by the baseline ViT encoder, the GTLI module, and the proposed MPGR module. Results on LUAD-HistoSeg. As illustrated in Fig. 2, the activation map produced by the ViT encoder (Fig. 2(b)) is relatively coarse and exhibits scattered responses, failing to capture precise structural boundaries. The GTLI module mainly improves localization completeness and suppresses noisy activations by injecting global semantic priors into patch representations. Under weak supervision, the CLS token provides image-level contextual guidance that helps recover incomplete activation regions and reduces irrelevant responses. In addition, the channel-wise gating mechanism stabilizes optimization by encouraging consistent global-local feature alignment during training. The GTLI module (Fig. 2(c)) improves the localization ability to some extent, but still suffers from fragmented activations and incomplete region coverage.Fig. 2Heatmap on LUAD-HistoSeg Heatmap on LUAD-HistoSeg In contrast, the proposed MPGR module (Fig. 2(d)) generates significantly smoother and more spatially coherent activation maps. It effectively suppresses irrelevant background regions while highlighting semantically meaningful tissue structures. The final prediction (Fig. 3(e)) demonstrates clear object boundaries and improved region consistency, indicating that the multi-scale graph reasoning successfully enhances contextual aggregation and spatial continuity.Fig. 3Heatmap on BCSS-WSSS Heatmap on BCSS-WSSS Results on BCSS-WSSS. Similar observations can be made from Fig. 3. The ViT encoder (Fig. 3(b)) produces noisy and dispersed responses, while GTLI (Fig. 3(c)) introduces slight improvements but still lacks global consistency. The MPGR module (Fig. 3(d)) significantly enhances the activation quality by producing more concentrated and structured responses. Compared to previous methods, it better captures long-range dependencies and reduces false activations. The final prediction (Fig. 3(e)) exhibits sharper boundaries and more accurate region delineation, especially in complex tissue areas. Overall, the visualization results clearly demonstrate the advantages of MPGR-Net:Enhanced spatial coherence: Multi-scale graph reasoning enforces consistent predictions across neighboring patches.Improved boundary delineation: The model produces smoother and more accurate segmentation boundaries.Noise suppression: Irrelevant activations are effectively reduced compared to baseline methods. Enhanced spatial coherence: Multi-scale graph reasoning enforces consistent predictions across neighboring patches. Enhanced spatial coherence: Multi-scale graph reasoning enforces consistent predictions across neighboring patches. Improved boundary delineation: The model produces smoother and more accurate segmentation boundaries. Improved boundary delineation: The model produces smoother and more accurate segmentation boundaries. Noise suppression: Irrelevant activations are effectively reduced compared to baseline methods. Noise suppression: Irrelevant activations are effectively reduced compared to baseline methods. These observations validate that the proposed multi-scale graph reasoning mechanism plays a critical role in refining token representations and improving weakly supervised segmentation performance. Discussion We further analyze the behavior of MPGR-Net by jointly considering the quantitative results in Table 3 and the qualitative visualizations in Figs. 2 and 3, with a focus on the intrinsic differences between LUAD-HistoSeg and BCSS-WSSS. LUAD-HistoSeg exhibits relatively clear tissue organization with well-defined structural patterns, such as glandular or duct-like boundaries. As shown in Table 3, the best performance is achieved with a moderate connectivity configuration (e.g., (4, 8, 12, 16)), suggesting that capturing local and mid-range spatial dependencies is sufficient for modeling such structured patterns. This observation is consistent with the visualization in Fig. 2, where MPGR produces smooth and coherent activation maps with well-preserved boundaries. The model effectively enhances region consistency while avoiding excessive diffusion across distinct tissue regions. In contrast, BCSS-WSSS presents more diffuse and heterogeneous structures, where discriminative cues are often localized at the cellular level, such as nuclei distribution and subtle texture variations. As reflected in Table 3, configurations with larger connectivity ranges (e.g., including 20 and 24 neighbors) yield better performance. This indicates that broader receptive fields are beneficial for aggregating dispersed contextual information. The visualization in Fig. 3 further supports this observation: MPGR generates more globally consistent activation patterns and reduces noisy responses compared to the baseline methods, enabling better localization in structurally complex regions. These results highlight that the effectiveness of multi-scale graph reasoning is closely related to the structural characteristics of the dataset. The preferred graph receptive field size is closely related to the morphological characteristics of the dataset. LUAD-HistoSeg exhibits relatively regular and spatially continuous tissue structures, where moderate connectivity is sufficient for capturing local contextual dependencies while preserving boundaries. In contrast, BCSS-WSSS contains more heterogeneous and diffuse tissue distributions with higher intra-class variability, making broader receptive fields more beneficial for aggregating long-range contextual information. Overall, MPGR-Net demonstrates strong adaptability across different pathological scenarios by flexibly integrating multi-scale spatial relationships, allowing it to balance local detail preservation and global context modeling. Conclusion In this paper, we proposed MPGR-Net, a novel multi-scale patch graph reasoning framework for weakly supervised histopathological image segmentation. Unlike conventional methods that rely heavily on class activation maps, our approach explicitly models spatial dependencies among patch tokens through a multi-scale graph structure, enabling more effective contextual aggregation and structural reasoning.Specifically, we introduced a multi-scale graph construction strategy that captures spatial relationships at different receptive fields, along with a deep graph reasoning module to progressively refine token representations. Extensive experiments on LUAD-HistoSeg and BCSS-WSSS demonstrate that MPGR-Net consistently outperforms existing weakly supervised segmentation methods. Both quantitative results and qualitative visualizations verify that the proposed method produces more accurate, coherent, and structurally consistent segmentation maps. In addition, comprehensive ablation studies confirm the effectiveness of each component, including connectivity design, graph depth, and attention configuration. Moreover, our analysis reveals that the proposed multi-scale graph reasoning mechanism exhibits strong adaptability across datasets with different structural characteristics, effectively balancing local detail preservation and global context modeling. In future work, we plan to explore more adaptive graph construction strategies, integrate domain-specific priors such as histological knowledge, and extend the framework to large-scale whole-slide image analysis and other weakly supervised medical imaging tasks.