PathRWKV: Enhancing Whole Slide Image Inference With Asymmetric Recurrent Modeling

Area of Science:

• Digital Pathology
• Computational Biology
• Artificial Intelligence in Medicine

Background:

• Whole Slide Imaging (WSI) is crucial for cancer diagnosis, analyzing cellular to tissue details.
• Direct WSI processing faces GPU memory limits, leading to Multiple Instance Learning (MIL) using tile partitioning.
• Existing two-stage MIL methods have limitations in training throughput, inference memory, overfitting, spatial integrity, and multi-scale feature modeling.

Purpose of the Study:

• Introduce PathRWKV, a novel State Space Model for efficient and robust WSI analysis.
• Address computational trade-offs, overfitting, spatial context disruption, and inadequate multi-scale feature modeling in MIL.
• Develop a scalable WSI analysis solution with broad application potential.

Main Methods:

• Propose an asymmetric structure with max pooling for parallelized training and O(1) memory inference.
• Implement random sampling and multi-task learning to mitigate overfitting on limited WSI datasets.
• Utilize 2D sinusoidal position encoding for spatial context and TimeMix/ChannelMix modules for multi-scale feature modeling.

Main Results:

• PathRWKV demonstrates superior performance compared to 11 state-of-the-art methods.
• Achieved significant improvements across 10 out of 11 tested WSI datasets.
• Validated on a large-scale dataset of 29,073 WSIs.

Conclusions:

• PathRWKV offers an efficient and robust solution for WSI analysis in digital pathology.
• The model effectively overcomes limitations of existing MIL frameworks.
• PathRWKV presents a scalable approach with strong potential for clinical application in cancer diagnosis.