MS-DCANet: A Novel Segmentation Network For Multi-Modality COVID-19 Medical Images.

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Computational Pathology

Background:

• The COVID-19 pandemic presents challenges in medical image analysis due to blurred boundaries and low contrast.
• Existing segmentation methods often increase complexity, hindering clinical application.
• A need exists for models balancing accuracy, computational complexity, and inference speed.

Purpose of the Study:

• To propose MS-DCANet, a symmetric automatic segmentation framework for COVID-19 medical images.
• To develop a model that effectively captures lesion complexity, background relationships, and global context.
• To achieve a balance between segmentation accuracy and computational efficiency for clinical translation.

Main Methods:

• Introduced a novel Tokenized MLP block with a shift-window mechanism for fusing local and global features.
• Employed Dual Channel blocks and a Res-ASPP block to enhance small target recognition.
• Developed a symmetric network architecture for improved boundary continuity and spatial understanding.

Main Results:

• MS-DCANet achieved state-of-the-art performance on multi-modality COVID-19 datasets.
• The model demonstrated strong generalization capabilities on ISIC 2018 and BAA datasets.
• Achieved high MIOU scores (e.g., 73.86% on COVID-19 X-ray, 97.26% on COVID-19 CT).

Conclusions:

• MS-DCANet offers an effective solution for automated COVID-19 medical image segmentation.
• The model successfully balances accuracy and computational complexity, facilitating clinical deployment.
• Potential to assist clinicians in automating the diagnosis of COVID-19 patients with varying symptoms.