Gastric cancer: Contrastive learning without ImageNet detects tumors on blades and outlines their stages without annotation

Driven by self-supervision on colored H&E cuts, the model reaches a AUC of 0.91 to distinguish tumour and normal tissue, then groups tumours into four pseudo-stages.

Gastric cancer is one of the leading causes of cancer mortality, and its histopathological images are so variable that they complicate diagnosis, both human and automated, especially when annotated data are lacking. Aleakhya Pandeti, Pranav M. Pawar, Mithun Mukherjee, and Raja Muthalagu propose a self-supervised contrastive learning framework, based on a ResNet50 encoder trained without ImageNet pre-training, to learn directly from hematoxylin-eosin-stained tissue fragments representations that are insensitive to staining variations.

Detection, then stratification

Transferred to a supervised classifier, these representations distinguish tumor tissue and normal tissue with 83.1% accuracy in testing, an AUC-ROC of 0.909 and an average accuracy of 0.906, results comparable to recent supervised approaches while the model uses much fewer labeled samples. A K-means grouping of tumor representations then reveals four well-separated groups, assimilated to pseudo-stages, which a random forest finds at 97%: the model therefore encodes morphological variations related to progression without ever having seen a stage annotation.

Grad-CAM maps accompany each prediction to make it interpretable, and the system is offered via a Streamlit clinical interface. However, these pseudo-stages remain a statistical construct that will have to be compared with real clinical stages.