A Survey on Graph-Based Deep Learning for Computational Histopathology

TL;DR

This survey addresses the shift from patch-based CNN analysis to graph-based deep learning for computational histopathology, arguing that entity-graphs—cell-, patch-, tissue-, and hierarchical graphs—better capture micro- and macro-architectural tissue context. It synthesizes graph representations, core GNN architectures, pooling and interpretability methods, and a broad range of applications across breast, colorectal, prostate, lung, skin, cervical, and renal cancers, highlighting strong performance and interpretability benefits. The review also identifies critical open challenges, including graph construction, integration of expert knowledge, training efficiency, and explainability, and outlines future directions for clinical adoption. Overall, graph-based DL offers a principled path to richer tissue modeling and potential improvements in diagnostic accuracy and clinician trust, with practical impact in precision pathology. $Z$-score-like insights and topological analyses are used to explain tissue-level decisions, illustrating the potential for interpretable, multiscale pathology reasoning. The field remains nascent, with opportunities to standardize graph generation, leverage multimodal data, and translate methods into routine clinical workflows.

Abstract

With the remarkable success of representation learning for prediction problems, we have witnessed a rapid expansion of the use of machine learning and deep learning for the analysis of digital pathology and biopsy image patches. However, learning over patch-wise features using convolutional neural networks limits the ability of the model to capture global contextual information and comprehensively model tissue composition. The phenotypical and topological distribution of constituent histological entities play a critical role in tissue diagnosis. As such, graph data representations and deep learning have attracted significant attention for encoding tissue representations, and capturing intra- and inter- entity level interactions. In this review, we provide a conceptual grounding for graph analytics in digital pathology, including entity-graph construction and graph architectures, and present their current success for tumor localization and classification, tumor invasion and staging, image retrieval, and survival prediction. We provide an overview of these methods in a systematic manner organized by the graph representation of the input image, scale, and organ on which they operate. We also outline the limitations of existing techniques, and suggest potential future research directions in this domain.

Figures (18)

• Figure 1: Traditional CNNs excel at modelling local relations in grid representation, where the topology of the neighborhood is constant (Left). GCNs can take into account different neighbouring relations (global relation) by going beyond the local pixel neighbourhoods used by convolutions. On a graph, the neighbours of a node are unordered and variable in size (Right).
• Figure 2: Top: Graph-based representation of images for relation-aware human-object interaction, image segmentation, and human pose estimation (left-to-right). Images adapted from qi2018learningchen2019graphli2020temporal. Bottom:A. Cell-graph representation for prostate cancer. B. Tissue-graph representation for colorectal cancer. C. Hierarchical cell-to-tissue graph representation for breast cancer. Images adapted from wang2020weaklylevy2020topologicalpati2020hact.
• Figure 3: Overview of a standard graph-based workflow in computational pathology. The WSI image is first transformed into one or more graphs. 1. The entities can be nuclei, patches or tissue regions. 2. Node features comprise handcrafted or deep learning features to characterize the entities. 3. The edges encode intrinsic relationships (spatial or semantic) among the entities. 4. Graph encoding and classification (node-level or graph-level prediction): the graph representation is processed using GNNs and its variants such as ChebNet, GCN, GraphSAGE, GAT, and GIN, including different graph pooling strategies (global or hierarchical pooling). 5. Graph interpretations: a set of GNN model interpretability tools such as graph attentions or post-hoc graph explainers (e.g. GNNExplainer and GraphGrad-CAM.)
• Figure 4: Representation of graph architectures for node-level classification. Recreated from wu2020comprehensive.
• Figure 5: Representation of graph models for graph-level classification. Recreated from wu2020comprehensive.