Learning Basal Cell Carcinoma Patterns by Fusing Pathologists’ WSI Navigations and Graph-Based Centrality Features

This article introduces a model that automatically detects suspicious basal cell carcinoma regions in whole slide images (WSI) by integrating nuclear architectural features and information captured from WSI pathologists’ navigations during diagnostic tasks. In such an approach, manual annotations are not needed since high-level expert knowledge is implicitly captured when the pathologist is exploring the WSI. The method was tested on a set of 10 cases of patients diagnosed with basal cell carcinoma using a leave-one-out cross-validation technique. At each iteration, a quadratic discriminant analysis classifier was trained to identify cancerous nuclei using architectural features of nuclei belonging to the regions that were highly or little visited by pathologists when rendering a diagnosis. Experimental results showed an average accuracy of 86% and an F-score of 76%, thereby demonstrating the potential of this approach to be included in actual clinical scenarios.