Using Variational Autoencoders for Out of Distribution Detection in Histological Multiple Instance Learning

Abstract

In the context of histological image classification, Multiple Instance Learning (MIL) methods only require labels at Whole Slide Image (WSI) level, effectively reducing the annotation bottleneck. However, for their deployment in real scenarios, they must be able to detect the presence of previously unseen tissues or artifacts, the so-called Out-of-Distribution (OOD) samples. This would allow Computer Assisted Diagnosis systems to flag samples for additional quality or content control. In this work, we propose an OOD-aware probabilistic deep MIL model that combines the latent representation from a variational autoencoder and an attention mechanism. At test time, the latent representations of the instances are used in the classification and OOD detection tasks. We also propose a deterministic version of the model that uses the reconstruction error as OOD score. Panda (prostate tissue) and Camelyon16 (lymph node tissue) are used as train/test in-distribution datasets, obtaining bag classification results competitive with current stateof-the-art models. OOD detection is evaluated performing two experiments for each in-distribution dataset. For Panda, Camelyon16 and ARTIF (prostate tissue contaminated with artifacts) are used as OOD datasets, obtaining 100% AUC in both cases. For Camelyon16, Panda and BCELL (lymph node tissue diagnosed with diffuse large B-cell lymphoma) are used as OOD datasets, obtaining AUCs of 100% and 97%, respectively. Experimental validation demonstrates the models’ strong classification performance and effective OOD slide detection, highlighting their clinical potential.