Multi-diagnosis multi-instance learning for auxiliary gene mutation diagnosis in whole slide images

Area of Science:

• Computational pathology
• Genomic medicine
• Artificial intelligence in healthcare

Background:

• Multi-Instance Learning (MIL) is crucial for whole slide image (WSI) analysis in pathology.
• Challenges persist in linking WSIs to genetic mutation data due to non-intuitive correlations.
• Traditional models struggle to capture key genetic mutation features from histological images.

Purpose of the Study:

• To develop a novel cancer-gene multi-instance learning approach for WSI-level genetic mutation diagnosis.
• To integrate multi-diagnosis data, including pixel-level cancer annotations and WSI-level genetic mutation annotations.
• To systematically learn and represent genetic mutations by leveraging correlations between cancerous and mutated regions.

Main Methods:

• Proposed a multi-instance learning framework integrating cancer and genetic mutation diagnostic data.
• Utilized pixel-level cancer annotations and WSI-level genetic mutation annotations as multi-diagnosis inputs.
• Employed focused sampling and lesion filtering to identify representative instances and analyze morphological/histological features of genetic mutations.

Main Results:

• Achieved advanced performance in gene mutation datasets for lung, bladder, and breast cancers.
• Generated visualizations for model interpretability, highlighting areas of attention.
• Demonstrated effective deep coupling of cross-diagnostic pathological representations.

Conclusions:

• The framework effectively deconstructs correlation maps between cancer and gene mutation regions in WSIs.
• Provides decision support for pathologists in clinical diagnosis.
• Shows potential utility for prioritizing cases in pathology workflows.