Deformable attention (DANet) for semantic image segmentation.

Area of Science:

• Medical image analysis
• Deep learning
• Computer vision

Background:

• Deep learning-based medical image segmentation is a key research area.
• Attention mechanisms enhance deep networks for semantic segmentation.
• Criss-cross attention offers efficiency but may miss pertinent non-local information.

Purpose of the Study:

• To propose a novel Deformable Attention Network (DANet) for more accurate medical image segmentation.
• To enhance contextual information computation in deep segmentation networks.
• To improve COVID-19 lesion segmentation accuracy.

Main Methods:

• Developed a new Deformable Attention Network (DANet) incorporating learnable feature map deformations.
• Applied DANet within a U-Net architecture for semantic segmentation.
• Evaluated DANet's performance on COVID-19 lesion segmentation tasks.

Main Results:

• DANet achieved a Dice score of 60.17% for COVID-19 lesion segmentation.
• The proposed network improved accuracy by 4.4 percentage points compared to a baseline U-Net.
• Recursive application of deformable attention blocks enhanced dynamic and precise attention context capture.

Conclusions:

• DANet enables more accurate contextual information computation in deep segmentation networks.
• The novel attention mechanism effectively captures relevant non-local features, boosting segmentation performance.
• DANet shows significant potential for improving medical image segmentation, particularly for tasks like COVID-19 lesion detection.