Rethinking Multiple Instance Learning for Whole Slide Image Classification: A Bag-Level Classifier is a Good...

Area of Science:

• Computational pathology
• Digital pathology
• Machine learning in medicine

Background:

• Multiple Instance Learning (MIL) shows potential for Whole Slide Image (WSI) classification.
• High computational cost of gigapixel WSIs is a major challenge.
• Existing two-stage MIL methods with fixed feature embedders lead to suboptimal accuracy due to stage disparity.

Purpose of the Study:

• To develop a cost-effective method for improving MIL-based WSI classification.
• To address the accuracy limitations caused by fixed feature embedders in existing MIL approaches.
• To couple the feature embedder and bag classifier iteratively for enhanced performance.

Main Methods:

• Proposed Iteratively Coupled Multiple Instance Learning (ICMIL) framework.
• Employs an iterative process: train bag classifier with fixed embedder, then fine-tune embedder with fixed classifier.
• Introduced a teacher-student framework for efficient knowledge distillation from bag classifier to instance-level embedder.

Main Results:

• ICMIL significantly improves the performance of existing MIL backbones.
• Achieved state-of-the-art results across four distinct WSI classification datasets.
• Demonstrated effectiveness in reducing computational cost while enhancing accuracy.

Conclusions:

• ICMIL offers an effective and low-cost solution for MIL-based WSI classification.
• The iterative coupling and knowledge distillation approach overcomes limitations of fixed embedders.
• The method provides a robust improvement over existing MIL techniques for digital pathology applications.