A Contrastive Learning Framework for Label-Efficient Gastric Cancer Detection and Intra-Tumour Progression Analysis

Abstract

Gastric Cancer is among the leading causes of cancer-related mortality worldwide. The complexity and variability of histopathological images pose significant challenges for both pathologists and computer-based diagnostics, creating a growing need for machine learning approaches capable of capturing meaningful representations from limited labelled data. Selfsupervised Contrastive Learning (CL) has demonstrated promising results in natural image analysis; however, its application to gastric cancer histopathology remains underexplored. This paper proposes an integrated CL framework built on a ResNet50 encoder trained under the NT-Xent objective, without reliance on ImageNet pre-training, to learn discriminative and staining-invariant representations from H&E-stained tissue patches. The learned embeddings are transferred to a supervised downstream classifier for binary tumour/normal (TUM/NOR) detection, achieving a test accuracy of 83.13%, an AUC-ROC of 0.9087, and an Average Precision of 0.9063 - competitive with recent supervised and weakly-supervised approaches on comparable histopathology benchmarks. Beyond binary diagnosis, the framework applies K-Means clustering to tumour embeddings to identify four geometrically separable pseudo-stage clusters, confirmed by a Random Forest classifier achieving 97% accuracy and a mean Dice coefficient of 0.97, demonstrating the capacity of CL representations to encode progression-relevant morphological variation without stage-level annotations. Grad-CAM explainability is integrated throughout to ensure model transparency. The complete system is deployed via a Streamlit clinical interface. Compared to recent state-of-the-art methods the proposed framework operates with substantially fewer labelled samples while additionally providing unsupervised progression stratification and interpretable outputs, addressing a gap not covered by existing approaches.