LVAERMIL: enhancing multiple instance learning robustness via latent variational autoencoders

Abstract Multiple Instance Learning (MIL) has become widely used in medical imaging classification, where precisely annotated data is scarce. MIL methods treat whole slide images as bags of instances (the patches), and only bag labels are required for training. Recently, MIL methods have shown extraordinary results in both classification and tumor localization benchmark tasks. However, in real-world scenarios, slides may be contaminated by tissue from other specimens or artifacts that can extrapolate MIL models and produce medical errors. This presents a challenge for widespread use of MIL in pathology labs. In this work, we propose Latent-VAE-Robust-MIL (LVAERMIL), a novel probabilistic module that learns the instance-level data distribution using a Variational Autoencoder trained on the latent space produced by a feature extractor. This enables the detection of out-of-distribution patches and their elimination from MIL bags, improving the robustness of any MIL model to whole slide image contaminants. The extensive evaluation performed by contaminating two datasets using four foreign tissues and using two scanner-induced artifacts shows that LVAERMIL maintains or improves the accuracy of the eight state-of-the-art MIL models on contaminated datasets, especially when the amount of contamination in the slides is high.