CoviSwin: A Deep Vision Transformer for Automatic Segmentation of COVID-19 CT Scans

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate segmentation of COVID-19 lesions in chest CT scans is crucial for patient care.
• Existing segmentation methods face challenges due to heterogeneous lesion appearances and limited labeled data.

Purpose of the Study:

• To introduce CoviSwin, a Transformer-based U-shaped network for enhanced COVID-19 lesion segmentation in CT scans.
• To address limitations of current methods in handling diverse lesion appearances and data scarcity.

Main Methods:

• Developed CoviSwin, a U-shaped encoder-decoder network integrating Swin Transformer V2, attention, and residual connections.
• Employed a two-phase training strategy using SemiSeg and MedSeg datasets.
• The network captures both global context and fine-grained details for precise segmentation.

Main Results:

• CoviSwin achieved a mean sensitivity of 0.790 ± 0.012 and a Dice Similarity Coefficient (DSC) of 0.781 ± 0.0068.
• The model demonstrated a specificity of 0.962 ± 0.0049.
• Outperformed recent models like NextSeg and GFNet in sensitivity by 8.07% and 7.48%, respectively.

Conclusions:

• CoviSwin shows significant potential as an effective tool for clinical COVID-19 lesion segmentation.
• The model's architecture and training strategy contribute to its superior performance.
• Highlights the advancement in AI-driven medical image analysis for infectious diseases.