Multi-Instance Multi-Label Learning for Multi-Class Classification of Whole Slide Breast Histopathology Images

Area of Science:

• Digital pathology
• Computational pathology
• Machine learning in histopathology

Background:

• Whole slide imaging presents challenges in correlating image regions with diagnostic labels.
• Existing algorithms struggle with multi-class classification and localization in digital pathology.
• Weakly supervised learning is needed to address these challenges.

Purpose of the Study:

• To develop and evaluate weakly supervised learning methods for multi-class classification and localization in digital pathology.
• To leverage pathologist viewing records for improved diagnostic accuracy.
• To address the limitations of current algorithms in whole slide image analysis.

Main Methods:

• Extracting candidate regions of interest (ROIs) from pathologist viewing logs (zooming, panning, fixation).
• Modeling whole slides as bags of instances using candidate ROIs and slide-level annotations.
• Applying multi-instance multi-label learning algorithms for slide-level and ROI-level predictions.

Main Results:

• Achieved slide-level average precision up to 81% (5-class) and 69% (14-class) in weakly labeled scenarios.
• Demonstrated successful multi-class localization and classification at the ROI level within whole slide images.
• Validated the approach on challenging diagnostic categories in breast histopathology.

Conclusions:

• Pathologist viewing records can be effectively utilized in weakly supervised learning for digital pathology.
• The proposed methods show promise for accurate multi-class classification and localization in whole slide images.
• This approach advances the capabilities of artificial intelligence in diagnostic pathology.