SI-MIL: Taming Deep MIL for Self-Interpretability in Gigapixel Histopathology.

Area of Science:

• Computational pathology
• Digital pathology
• Machine learning for medical imaging

Background:

• Interpreting Multiple Instance Learning (MIL) for Whole Slide Image (WSI) analysis is complex due to gigapixel slide size.
• Current MIL interpretability methods offer limited insight into the rationale behind region selection for pathologists.

Purpose of the Study:

• To develop a Self-Interpretable Multiple Instance Learning (SI-MIL) method for WSI analysis that intrinsically incorporates interpretability.
• To provide feature-level interpretations grounded in pathological insights, going beyond salient region identification.

Main Methods:

• A deep MIL framework guiding an interpretable branch based on handcrafted pathological features for linear predictions.
• Challenging the performance-interpretability trade-off using linear prediction constraints within the SI-MIL framework.

Main Results:

• SI-MIL demonstrates competitive WSI-level prediction performance across three cancer types compared to state-of-the-art methods.
• The method provides unique feature-level interpretations rooted in pathological insights for WSIs.
• Comprehensive benchmarking validates SI-MIL's local and global interpretability, user-friendliness, and faithfulness.

Conclusions:

• SI-MIL offers a novel approach to interpretable MIL for WSI analysis, providing both accurate predictions and actionable pathological insights.
• The findings challenge the notion that improved interpretability necessitates a compromise in predictive performance for complex medical imaging tasks.