SIB-MIL: Sparsity-Induced Bayesian Neural Network for Robust Multiple Instance Learning on Whole Slide Image Analysis

Area of Science:

• Computational pathology
• Machine learning in histopathology
• Medical image analysis

Background:

• Multiple instance learning (MIL) is effective for whole slide histopathology images (WSIs) but struggles with overfitting and uncertainty quantification.
• Existing Bayesian neural networks (BNNs) face challenges with unstable predictions and high variance under weak supervision in WSI analysis.

Purpose of the Study:

• To develop a robust MIL method for WSI analysis that mitigates overfitting and provides reliable uncertainty quantification.
• To address the limitations of Gaussian BNNs in WSI prediction by introducing a sparsity-induced prior.

Main Methods:

• Proposed SIB-MIL: a sparsity-induced Bayesian Neural Network integrated into the MIL framework.
• Utilized a Horse-shoe prior on BNN parameters to induce sparsity, filter noise, and manage prediction variance.
• Applied the method to cancer classification and subtyping tasks using WSIs.

Main Results:

• SIB-MIL demonstrated improved performance over existing MIL networks in WSI analysis.
• The method effectively addressed variance overflowing issues common in Gaussian BNNs.
• Achieved robust performance in uncertainty quantification for histopathology image tasks.

Conclusions:

• SIB-MIL offers a more robust and uncertainty-aware approach for analyzing whole slide histopathology images.
• The sparsity-induced prior is crucial for enhancing MIL performance and reliability in WSI analysis.
• This work advances computational pathology by providing a powerful tool for cancer diagnosis and subtyping.