Few-shot learning and explainable AI for colon cancer histopathology: A prototypical network with multi-technique interpretability

Abstract

Background Colon cancer diagnosis from histopathology is challenging due to limited annotated data and the lack of interpretability in deep models. Objective We present a data-efficient framework combining few-shot learning and explainable AI for accurate and transparent diagnosis. Methods A Prototypical Network with a ConvNeXt-Tiny backbone was trained on small colon-tissue image sets. Explanations from Grad-CAM and LIME were validated by a pathologist, and generalization was tested on an external dataset. Results The model achieved 98.5 % accuracy in-domain and 90 % on the EBHI dataset, showing strong generalization. Conclusions This few-shot and explainable model performs well with minimal data and generates clinically interpretable visual outputs, supporting its potential for reliable colon cancer diagnostics. • Few-shot Prototypical Network classifies colon H&E slides with up to 98.5 % accuracy. • Multi-technique XAI (Grad-CAM & LIME) confirmed by a board-certified pathologist. • 90 % cross-domain accuracy on EBHI shows robustness to scanner and stain shifts.