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

Area of Science:

• Computational pathology
• Machine learning for medical imaging
• Histopathology analysis

Background:

• Multiple Instance Learning (MIL) methods simplify histological image classification by using Whole Slide Image (WSI) level labels, reducing annotation burden.
• Real-world deployment requires MIL models to identify Out-of-Distribution (OOD) samples, such as novel tissues or artifacts, for quality control in Computer Assisted Diagnosis (CADx).

Purpose of the Study:

• To develop an OOD-aware probabilistic deep MIL model for histological image classification.
• To enable CADx systems to flag potentially problematic samples for further review.

Main Methods:

• A novel OOD-aware probabilistic deep MIL model integrating a variational autoencoder's latent representation with an attention mechanism was proposed.
• The model utilizes instance latent representations for both classification and OOD detection at test time.
• A deterministic variant using reconstruction error as an OOD score was also developed.

Main Results:

• The model achieved classification results competitive with state-of-the-art methods on Panda (prostate) and Camelyon16 (lymph node) datasets.
• For OOD detection, the model achieved 100% AUC when distinguishing between prostate tissue and artifacts (artif dataset).
• The model also demonstrated high AUCs (100% and 97%) for lymph node OOD detection using Panda and B-cell lymphoma (bcell) datasets.

Conclusions:

• The developed models exhibit strong classification performance and effective OOD slide detection capabilities.
• These findings highlight the clinical potential of the proposed OOD-aware MIL approach for enhancing diagnostic accuracy and reliability.