[Paper Review] Context-aware Skin Cancer Epithelial Cell Classification with Scalable Graph Transformers

Lucas Sancéré, Noémie Moreau+1 more|arXiv (Cornell University)|Feb 17, 2026

Cutaneous Melanoma Detection and Management0 citations

TL;DR

The paper demonstrates scalable Graph Transformers on full-WSI cell graphs to classify healthy versus tumor epithelial cells in cutaneous squamous cell carcinoma, outperforming patch-based image methods and offering faster training times.

ABSTRACT

Whole-slide images (WSIs) from cancer patients contain rich information that can be used for medical diagnosis or to follow treatment progress. To automate their analysis, numerous deep learning methods based on convolutional neural networks and Vision Transformers have been developed and have achieved strong performance in segmentation and classification tasks. However, due to the large size and complex cellular organization of WSIs, these models rely on patch-based representations, losing vital tissue-level context. We propose using scalable Graph Transformers on a full-WSI cell graph for classification. We evaluate this methodology on a challenging task: the classification of healthy versus tumor epithelial cells in cutaneous squamous cell carcinoma (cSCC), where both cell types exhibit very similar morphologies and are therefore difficult to differentiate for image-based approaches. We first compared image-based and graph-based methods on a single WSI. Graph Transformer models SGFormer and DIFFormer achieved balanced accuracies of $85.2 \pm 1.5$ ($\pm$ standard error) and $85.1 \pm 2.5$ in 3-fold cross-validation, respectively, whereas the best image-based method reached $81.2 \pm 3.0$. By evaluating several node feature configurations, we found that the most informative representation combined morphological and texture features as well as the cell classes of non-epithelial cells, highlighting the importance of the surrounding cellular context. We then extended our work to train on several WSIs from several patients. To address the computational constraints of image-based models, we extracted four $2560 imes 2560$ pixel patches from each image and converted them into graphs. In this setting, DIFFormer achieved a balanced accuracy of $83.6 \pm 1.9$ (3-fold cross-validation), while the state-of-the-art image-based model CellViT256 reached $78.1 \pm 0.5$.

Motivation & Objective

• Address the limitation of patch-based WSIs that lose tissue-level context during analysis.
• Propose a full-WSI cell-graph representation with scalable Graph Transformers to classify epithelial cells.
• Systematically compare graph-based methods against image-based approaches on WSI-Graph and TILE-Graphs datasets.
• Investigate how node feature choices and graph simplification affect classification performance.

Proposed method

• Construct a WSI-level cell graph where nodes are nuclei with morphological, texture, and class features, and edges connect neighboring nuclei within a threshold distance.
• Refine epithelial cell labels using expert tumor annotations to create tumor vs healthy epithelial node classes.
• Simplify graphs by retaining nodes within k-max hops of epithelial anchors to balance context and computational efficiency.
• Evaluate linear-complexity Graph Transformers (SGFormer, NodeFormer, DIFFormer) for binary node classification (tumor vs healthy) with masked target-class features to preserve contextual information.
• Compare graph-based models to image-based baselines (CellViT256) on WSI-Graph and TILE-Graphs across folds, using 3-fold cross-validation without early stopping.
• Train with Adam on large GPUs, adapting hyperparameters from prior benchmarks for each model; perform both subgraph and random-node evaluation protocols to gauge generalization.

Experimental results

Research questions

• RQ1 Can a full-WSI cell-graph representation improve discrimination between healthy and tumor epithelial cells compared to patch-based image methods?
• RQ2 Which node features (morphology, texture, cell class) and normalization strategies most improve epithelial cell classification?
• RQ3 How do scalable Graph Transformers with linear complexity perform versus traditional GNNs and image-based models on WSI-Graph and TILE-Graphs datasets?
• RQ4 What is the impact of graph simplification (max-hop) on classification accuracy and robustness across evaluation protocols?

Key findings

• On a single WSI, SGFormer achieved 85.2 ± 1.5 balanced accuracy and DIFFormer 85.1 ± 2.5, outperforming the best image-based method at 81.2 ± 3.0.
• On TILE-Graphs that span multiple patients, DIFFormer reached 83.6 ± 1.9 balanced accuracy versus CellViT256 at 78.1 ± 0.5.
• Graph-based training was substantially faster, with DIFFormer training taking about 32 minutes per fold versus ~5 days for CellViT256.
• Node feature ablation shows the combination of morphology, texture, and cell-class features with z-score normalization yields the best generalization (e.g., morphology+texture+cell class with normalization: 84.0 ± 2.8 subgraphs; 94.5 ± 0.4 random nodes; table data cited).
• Graph simplification with 10 max-hops provides a balance between connectivity and performance (86.6 ± 2.2 subgraphs; 95.0 ± 0.2 random nodes).
• Overall, graph-based approaches (DIFFormer, SGFormer) can outperform image-based methods on both within-patient and multi-patient datasets while offering substantial computational efficiency.