Self-Supervision Revives Simple Multiple Instance Classification Methods in Pathology

Ali Mammadov, Loic Le Folgoc, Julien Adam, Anne Buronfosse, Gilles Hayem, Guillaume HOCQUET, Pietro Gori

Keywords: Self-Supervised Learning, Multiple Instance Learning, Digital Pathology

Abstract: Multiple Instance Learning (MIL) is the current solution for classifying whole slide pathology images (WSI). MIL divides WSIs into patches, treating each slide as a bag of instances with a global label. There are two main MIL approaches: instance-based and embedding-based. The former classifies patches independently and aggregates scores for bag label prediction, while the latter performs bag classification after aggregating patch embeddings. Even if instance-based methods are more interpretable, embedding-based MILs have been preferred in the past, due to their robustness to poor feature extractors. In parallel, many works started to use self-supervised learning (SSL) for training better encoders. However, despite the use of SSL feature extractors, many works continue to endorse the superiority of embedding-based MILs. Here, we show that with a good SSL feature extractor, simple instance-based MILs, with very few parameters, obtain similar or better performance than complex, state-of-the-art embedding-based MIL methods.

Submission Number: 84