HMIL: Hierarchical Multi-Instance Learning for Fine-Grained Whole Slide Image Classification.

Area of Science:

• Computational pathology
• Digital pathology
• Precision oncology

Background:

• Fine-grained classification of whole slide images (WSIs) is crucial for precision oncology, demanding the identification of subtle morphological differences.
• Existing multi-instance learning (MIL) methods often fail to leverage hierarchical label structures, treating classification as a flat problem.

Purpose of the Study:

• To introduce a novel Hierarchical Multi-Instance Learning (HMIL) framework to address the limitations of current MIL approaches in fine-grained WSI classification.
• To improve the structured learning process by aligning inherent relationships between hierarchical labels at both instance and bag levels.

Main Methods:

• Developed a Hierarchical Multi-Instance Learning (HMIL) framework incorporating a class-wise attention mechanism for hierarchical information alignment.
• Integrated supervised contrastive learning to boost discriminative capabilities for fine-grained classification.
• Implemented a curriculum-based dynamic weighting module for adaptive balancing of hierarchical features during training.

Main Results:

• Demonstrated state-of-the-art performance on large-scale cytology cervical cancer (CCC) and public histology datasets (BRACS, PANDA).
• Achieved superior class-wise and overall classification accuracy compared to existing methods.
• Validated the effectiveness of the HMIL framework in capturing hierarchical label correlations.

Conclusions:

• The proposed HMIL framework offers a significant advancement in fine-grained classification of WSIs by effectively utilizing hierarchical label information.
• HMIL provides a more structured and informative approach, leading to improved diagnostic accuracy in precision oncology.
• The framework's performance highlights its potential for enhancing cancer diagnosis and personalized treatment strategies.