Can Graph Reasoning Improve Contextual Understanding in Histopathology?

Background: Automated classification of histopathology tissue types is a fundamental task in computational pathology, yet standard deep learning architectures process image patches independently, potentially neglecting relational structure between tissue regions. Objectives: In this work, we rigorously evaluate whether augmenting a ResNet-18 backbone with a Graph Convolutional Network (GCN) that performs message passing over a feature-space k-nearest-neighbor graph, built from patch feature similarity rather than true spatial coordinates, improves patch-level tissue classification. Methods: In an earlier version of this evaluation, we found that near-perfect validation accuracy was in fact an artifact of label leakage: the graph was constructed using ground-truth labels at both training and evaluation time, which cannot be reproduced at inference. We correct this throughout, using a label-free graph construction, an independent external test set (CRC-VAL-HE-7K, a disjoint patient cohort), 5-fold cross-validation, and two additional models: a Graph Attention Network (GAT) variant and a non-graph attention-pooling baseline. Because cross-validation folds share overlapping training data and a common fixed test set, we additionally report a corrected resampled significance test that accounts for this non-independence, rather than relying solely on a naive paired t-test. Results: Under this corrected protocol on NCT-CRC-HE-100K, none of the three graph-based variants differs significantly from the backbone-only model (95.48% ± 0.36% accuracy) on either accuracy or macro-F1 (corrected p ≥ 0.25 throughout), though all three point estimates are at or below the backbone. On a second, independent public dataset (Kather-2016), the GCN variant shows a small but statistically significant improvement that survives the corrected test (+0.48 points accuracy, corrected p = 0.0034), while GAT and the attention-pooling baseline show no effect, an inconsistent, dataset-dependent pattern rather than a reliable general benefit. We also directly quantify the leakage artifact across all 5 folds on Kather-2016, reproducing the original protocol to show it inflates accuracy by roughly 5-6 percentage points over the corrected result, consistently across folds. Conclusion: our results indicate that, once correctly evaluated, this form of feature-space graph-based relational reasoning provides at best a small, dataset-dependent benefit, not the large, consistent improvement suggested by the original, leaky evaluation, and we report the leakage mechanism and evaluation pitfalls in detail as a cautionary, reproducible example for related work.