Slide-Based Graph Collaborative Training for Histopathology Whole Slide Image Analysis.

Area of Science:

• Computational pathology
• Digital pathology
• Machine learning in oncology

Background:

• Pathological characteristics in whole slide images (WSIs) are crucial for cancer diagnostics.
• Current WSI analysis often overlooks inter-slide correlations, missing vital information from cancer development processes.
• Tumor development involves continuous histological, morphological, and genetic changes across different stages and patients.

Purpose of the Study:

• To introduce a novel computational pathology pipeline, SlideGCD, for enhanced WSI representation learning.
• To incorporate inter-slide correlations into WSI analysis to improve diagnostic accuracy.
• To adapt SlideGCD to existing Multiple Instance Learning (MIL) frameworks.

Main Methods:

• Proposed a generic WSI analysis pipeline named SlideGCD.
• Integrated prior knowledge of cancer development into an end-to-end workflow.
• Utilized slide-based graphs for guided message passing and refined slide representation.
• Adapted SlideGCD to 8 state-of-the-art MIL frameworks.

Main Results:

• Demonstrated the effectiveness and robustness of SlideGCD across 4 diverse tasks: cancer subtyping, staging, survival prediction, and gene mutation prediction.
• Validated improvements in WSI analysis by incorporating inter-slide correlations.
• Showcased the pipeline's adaptability with various MIL backbones.

Conclusions:

• SlideGCD offers a significant advancement in WSI analysis by effectively modeling inter-slide relationships.
• The pipeline enhances the performance of existing MIL frameworks for computational pathology tasks.
• Incorporating cancer development knowledge through inter-slide correlations improves WSI representation learning and diagnostic guidance.