PROTOTYPE GOVERNED CONTRASTIVE LEARNING FOR ROBUST IMAGE CLASSIFICATION IN HISTOPATHOLOGY

When covariate and domain shifts are present, the performance of deep learning models suffers dramatically. In a safety-critical discipline like histopathology, the management of out-of-distribution (OOD) samples remains a significant obstacle. This work proposes a supervised training and prediction technique using the innovative Prototype-Governed Contrastive Loss (PGCL) to solve limitations of standard classification methods in dealing with OOD samples. We demonstrate that the proposed approach improves OOD detection without sacrificing in-distribution (ID) classification performance. Extensive tests are conducted on a dataset of human colorectal cancer cases utilizing different CNN and Transformer-based model architectures. Several OOD detection methods described in the scientific literature are also compared. The prediction confidence derived from the PGCL framework resulted in a considerable performance increase over methods using cross-entropy models and comparable performance to semi-supervised approaches.