A Dual-branch Network with Cross-scale Feature Interaction and Alignment for Weakly Supervised Whole Slide Image...

Area of Science:

• Computational pathology
• Digital pathology
• Medical image analysis

Background:

• Whole slide images (WSIs) analysis is crucial for computer-aided diagnosis.
• Weakly supervised multiple instance learning (MIL) is widely used for WSI processing due to the lack of pixel-level annotations.
• Existing MIL methods often fail to fully utilize the pyramidal structure of WSIs and capture inter-instance dependencies and local context.

Purpose of the Study:

• To propose FIA-MIL, a dual-branch network for weakly supervised WSI analysis.
• To address limitations in capturing hierarchical information and inter-instance dependencies in current MIL approaches.
• To improve the performance of WSI analysis by incorporating cross-scale feature interaction and alignment.

Main Methods:

• Developed FIA-MIL, a dual-branch network integrating cross-scale feature interaction and alignment.
• Employed a dual-scale feature interaction module with Transformer encoders to model semantic relationships across magnifications and capture instance-level dependencies.
• Utilized a dual-scale feature aggregation module with alignment constraints for multi-scale feature integration and semantic consistency.

Main Results:

• The proposed FIA-MIL method demonstrated promising performance in classification and survival analysis tasks.
• Experiments were conducted on publicly available whole slide images datasets.
• The method effectively leverages the pyramidal structure and captures hierarchical information within WSIs.

Conclusions:

• FIA-MIL offers an effective approach for weakly supervised whole slide images analysis.
• The dual-branch network architecture with cross-scale feature interaction and alignment improves the capture of hierarchical information and dependencies.
• The proposed method shows potential for advancing computer-aided diagnosis in digital pathology.