Domain-Specific Transfer Learning for Gastric Cancer Tissue Classification

Gastric cancer is the fifth most diagnosed cancer worldwide and is the fifth most deadly. An essential part of the diagnosis and prognosis of cancer is the histopathological tissue classification. Pathologists classify tissue samples, but this is labour-intensive and can produce inconsistent results. In this study, we analyse the performance of CTransPath, a Swin Transformer model, for eight-class tissue classification after being pretrained on 15 million histopathological images, using the HMU-GC-HE-30K dataset, which contains 31,096 samples of gastric tissues. In a tenfold stratified cross-validation, we use a two-stage transfer learning model with a custom combination of histopathology-related data augmentation, class-weighted loss, and cosine annealing. The model reaches a macro AUC of 96.82% (±0.23%) (bootstrap 95% CI 96.70-96.91%) with patch-level accuracy of 76.74% (95% CI 76.27-77.20%) and macro F1 of 76.73% (95% CI 76.27-77.17%), indicating strong patch-level discriminative ability across the eight tissue classes in this internal single-institution evaluation. A dedicated calibration analysis (Brier score 0.325, expected calibration error 0.027 at 15 bins, calibration slope 1.14) shows that the predicted probabilities are reasonably well calibrated, with a mild tendency towards over-confidence. Additionally, a 4-variant ablation study shows that unfreezing the backbones provides a 3.4 to 3.6 percentage-point improvement over fixed-backbone models. Compared to multiple data models and architectures including the recently published pathology models UNI (ViT-L, 100M+ patches) and Virchow2 (ViT-H, 3.1M slides), CTransPath has the least parameters and provides the best performance within this controlled internal patch-level comparison. Under the same threefold protocol, CTransPath retained a statistically significant patch-level accuracy advantage over ImageNet-pretrained Swin-T and ResNet-50 (paired bootstrap $p<0.001$ ), as well as over the much larger UNI and Virchow2 foundation models (paired bootstrap $p<0.001$ and $p=0.002$ , respectively). Task-oriented model design and domain-specific pretraining appeared at least as effective as simply increasing model size for this task. Additionally, Grad-CAM, SHAP, t-SNE, and UMAP analyses provide qualitative evidence that the model attends to histologically meaningful regions as qualitative interpretability checks, not as confirmation of clinically validated reasoning. Because the public dataset lacks slide- and patient-level identifiers, these results should be interpreted strictly as single-institution, patch-level internal validation; potential data leakage from visually related patches cannot be excluded, and external multi-institution validation with slide-/patient-level evaluation is required before any clinical use. Our methodology and results provide a reproducible research benchmark for computational pathology research focused on gastric cancer.