Multiresolution Aggregation Transformer UNet Based on Multiscale Input and Coordinate Attention for Medical Image...

Area of Science:

• Medical Image Analysis
• Artificial Intelligence in Medicine

Background:

• UNet and transformer architectures are state-of-the-art for medical image segmentation.
• Multiscale feature fusion is critical for improving segmentation accuracy.
• Existing transformer-based UNet methods have limitations in exploring multiscale feature fusion.

Purpose of the Study:

• To propose a novel multiresolution aggregation transformer UNet (MRA-TUNet) for improved medical image segmentation.
• To enhance multiscale feature fusion using multiresolution input and coordinate attention.
• To achieve superior segmentation performance compared to existing methods.

Main Methods:

• Developed a multiresolution aggregation module for fusing input image information at different resolutions.
• Implemented an output feature selection module to integrate features from various scales.
• Introduced coordinate attention to further boost segmentation performance.
• Utilized multiscale input and coordinate attention for multiresolution aggregation.

Main Results:

• Achieved an average Dice score of 0.911 for right ventricle (RV), 0.890 for myocardium (Myo), 0.961 for left ventricle (LV), and 0.923 for left atrium (LA).
• Outperformed eight state-of-the-art methods in Dice score, precision, and recall on two benchmark datasets.
• Demonstrated superior performance in medical image segmentation tasks.

Conclusions:

• The proposed MRA-TUNet effectively enhances multiscale feature fusion for medical image segmentation.
• The integration of multiresolution aggregation and coordinate attention leads to significant performance improvements.
• MRA-TUNet represents a promising advancement in automated medical image segmentation.