Ranking-Aware Multiple Instance Learning for Histopathology Slide Classification: Development and Validation Study

Area of Science:

• Digital pathology
• Computational pathology
• Machine learning in medicine

Background:

• Multiple instance learning (MIL) is a key technique for slide-level classification in digital pathology.
• Current MIL methods often do not leverage partial expert annotations effectively.
• Expert annotations, even if partial, can significantly enhance supervised learning models.

Purpose of the Study:

• To develop and evaluate a ranking-aware MIL framework, named rank induction.
• To integrate partial expert annotations into MIL for improved slide-level classification.
• To assess the framework's performance under realistic annotation constraints.

Main Methods:

• Developed rank induction, a MIL approach utilizing a pairwise rank loss inspired by RankNet.
• The framework prioritizes diagnostically relevant patches by assigning higher attention to annotated regions.
• Evaluated on Camelyon16, DigestPath2019, and SMF-stomach datasets under various annotation scenarios.

Main Results:

• Rank induction achieved high AUROC scores: 0.839 (Camelyon16), 0.995 (DigestPath2019), and 0.875 (SMF-stomach).
• The model demonstrated robustness in low-data regimes, maintaining 0.761 AUROC with reduced training data.
• Near-saturated performance was achieved with only 20% sparse slide-level annotations.

Conclusions:

• Integrating expert annotations via ranking-based supervision enhances MIL-based classification performance.
• Rank induction proves practical and robust for digital pathology applications with limited, coarse, or sparse annotations.