PMSFINet: Progressive Multi-Scale Feature Interaction Network for Medical Image Segmentation.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Deep Learning

Background:

• Swin Transformer excels in dense prediction tasks by reducing computational complexity via window-based multi-head self-attention.
• Limitations exist in Swin Transformer for medical image segmentation, specifically in multi-scale feature fusion and boundary preservation of complex structures.

Purpose of the Study:

• To propose PMSFINet, a novel medical image segmentation network.
• To enhance representation learning through progressive multi-scale feature interaction for improved segmentation accuracy.

Main Methods:

• Developed a Progressive Multi-Scale Feature Interactive (PMSFI) module with Dual-Scale Window Interactive Attention (DSWIA) blocks for efficient computation and cross-scale information exchange.
• Integrated a Multi-Scale Super-Resolution Decoder (MSRD) with super-resolution, spatial attention, and a Local Similarity-Aware Sampler (LSAS) for refining details and enhancing boundaries.
• Employed a Cross-Attention Fusion (CAF) module with hybrid attention for dynamic fusion of dual-branch features, improving feature complementarity.

Main Results:

• Achieved Dice scores of 84.94% on Synapse, 92.43% on ACDC, and 90.79% on ISIC2018 datasets.
• Demonstrated strong generalization and robustness across diverse medical imaging tasks.
• Ablation studies confirmed the effectiveness of individual proposed components.

Conclusions:

• PMSFINet effectively addresses limitations in multi-scale feature fusion and boundary preservation for medical image segmentation.
• The proposed network shows significant potential for improving the accuracy and reliability of automated medical image analysis.
• PMSFINet offers a robust solution for segmenting complex and ambiguous structures in various medical imaging modalities.