SELDNet: Sequenced encoder and lightweight decoder network for COVID-19 infection region segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pulmonology

Background:

• Accurate segmentation of lung infections in computed tomography (CT) images is crucial for quantifying Coronavirus disease 2019 (COVID-19) severity and guiding treatment.
• Challenges in COVID-19 CT image segmentation include complex and variable imaging features, and high similarity to other lung pathologies, hindering accurate feature extraction.

Purpose of the Study:

• To address the challenges in segmenting COVID-19 lung infections from CT images.
• To introduce a novel deep learning model, SELDNet, for enhanced medical image segmentation.

Main Methods:

• Developed SELDNet, a sequence encoder and lightweight decoder network utilizing Transformer and deep separable convolution for fine-grained feature extraction.
• Integrated a semantic association module with a cross-attention mechanism to improve the fusion of multi-level semantic information between the encoder and decoder.

Main Results:

• SELDNet demonstrated effective segmentation of COVID-19 infected lung regions.
• Achieved a Dice score of 79.1%, sensitivity of 76.3%, and specificity of 96.7% in segmentation tasks.
• Outperformed several state-of-the-art image segmentation models in segmenting COVID-19 infected regions.

Conclusions:

• The proposed SELDNet model shows significant potential for accurate and efficient segmentation of COVID-19 lung infections in CT imaging.
• SELDNet offers a promising approach to overcome existing challenges in medical image segmentation for infectious lung diseases.
• The model's performance suggests its utility in clinical settings for disease monitoring and management.