Explainable CNN with spatial attention for oral cancer patch selection using a custom scoring metric

Objectives This study aims to propose an algorithm to select feature-rich patches from histopathology images, eliminating irrelevant regions, to support the computer-aided diagnosis of oral cancer. Methods To identify interpretable and diverse feature-rich patches, Haralick and local binary pattern (LBP) features are extracted and cosine similarity is calculated between patches. A trained convolutional neural network (CNN) model with spatial attention, along with a SHapley Additive exPlanation (SHAP) explainer, is used to estimate each patch’s contribution using SHAP values. These values, combined with cosine similarity, guide the selection of four patches that significantly influence the diagnosis while ensuring diversity. Features from the selected patches are extracted using a ResNet-50 model, concatenated, and normalized using Z-score normalization. The resulting feature vector is used to classify images as normal or cancerous. Results The approach was tested on a public dataset of 1,224 oral histopathology images from 230 patients, covering 100× and 400× magnifications. It achieved 97.53 % accuracy, 97.68 % precision, and 97.36 % sensitivity using a voting classifier, and 97.93 %, 99.43 %, and 97.76 % respectively with a stacking classifier, outperforming existing cancer detection methods. Conclusions The proposed algorithm effectively classifies oral biopsy images with minimal training, enhancing diagnostic reliability and offering valuable support for pathologists.