Patch-level phenotype identification via weakly supervised neuron selection in sparse autoencoders for CLIP-derived pathology embeddings

Computer-aided analysis of whole slide images (WSIs) has advanced rapidly with the emergence of multi-modal pathology foundation models. In this study, we propose a weakly supervised neuron selection approach to extract disentangled representations from CLIPderived pathology foundation models, leveraging the interpretability of sparse autoencoders. Specifically, neurons are ordered and selected using whole-slide level labels within a multiple instance learning (MIL) framework. We investigate the impact of different pre-trained image embeddings derived from general and pathology images and demonstrate that a selected single neuron can effectively enable patch-level phenotype identification. Experiments on the Camelyon16 and PANDA datasets demonstrate both the effectiveness and explainability of the proposed method, as well as its generalization ability for tumor patch identification.