DSKCA-UNet: Dynamic selective kernel channel attention for medical image segmentation

Area of Science:

• Medical Image Analysis
• Deep Learning for Medical Imaging
• Computer Vision in Healthcare

Background:

• U-Net is popular for medical image segmentation but struggles with remote dependencies.
• Transformers capture remote dependencies via self-attention (SA), but high dimensionality causes computational complexity.
• Existing SA methods overlook correlations between samples, limiting performance.

Purpose of the Study:

• To enhance Swin-UNet for improved medical image segmentation.
• To address limitations of self-attention in high-dimensional medical data.
• To develop a model with reduced complexity and robust performance.

Main Methods:

• Introduced a dynamic selective attention mechanism for convolution kernels in Swin-UNet.
• Implemented a local cross-channel interaction strategy without dimensionality reduction.
• Utilized global interaction between encoder features and a mixed loss function (weighted cross-entropy and Dice loss).

Main Results:

• Achieved a Dice similarity coefficient of 80.30% and 95% Hausdorff distance of 14.55% on the Synapse abdominal dataset.
• Obtained a Dice similarity coefficient of 90.80% on the ACDC cardiac dataset.
• Demonstrated good generalization ability and robustness on evaluated datasets.

Conclusions:

• The proposed dynamic selective attention mechanism improves Swin-UNet for medical image segmentation.
• The method effectively handles high-dimensional data, reduces model complexity, and alleviates category imbalance.
• The enhanced model shows significant potential as a powerful tool for medical image segmentation.