Dynamic deformable attention network (DDANet) for COVID-19 lesions semantic segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate medical image segmentation is crucial for disease diagnosis, particularly for COVID-19 CT scans where labeled data is scarce.
• Attention mechanisms in deep learning enhance contextual information capture for semantic segmentation tasks.
• Existing criss-cross attention modules offer efficiency but can be limited in accuracy by considering all non-local locations.

Purpose of the Study:

• To develop a novel Dynamic Deformable Attention Network (DDANet) for more accurate and efficient medical image segmentation.
• To improve the capture of spatial context in deep learning models for segmenting COVID-19 infection regions.
• To address the limitations of current attention mechanisms in handling scarce training data.

Main Methods:

• Proposed a Dynamic Deformable Attention Network (DDANet) incorporating a deformable criss-cross attention block.
• The attention block learns both attention coefficients and offsets continuously for precise contextual information.
• Integrated DDANet into a U-Net architecture for COVID-19 lesion segmentation.

Main Results:

• DDANet achieved Dice scores of 73.4% for Ground-glass opacity and 61.3% for consolidation lesions in COVID-19 segmentation.
• The proposed method demonstrated a 4.9% accuracy improvement over a baseline U-Net.
• DDANet showed a significant 24.4% improvement compared to current state-of-the-art methods.

Conclusions:

• The recursively applied dynamic deformable attention blocks effectively capture dynamic and precise attention context.
• DDANet offers a more accurate and efficient approach to medical image segmentation, especially for challenging tasks like COVID-19 lesion identification.
• The findings highlight the potential of DDANet in improving diagnostic accuracy with limited labeled data.