[Corona virus disease 2019 lesion segmentation network based on an adaptive joint loss function]

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Context:

• Accurate segmentation of COVID-19 lesions in CT images is crucial for diagnosis and treatment.
• Challenges include complex symptoms and blurred lesion boundaries, leading to potential misdiagnosis.

Purpose:

• To develop an advanced deep learning model for precise COVID-19 lesion segmentation.
• To introduce the Level Set Generalized Dice Loss (LGDL) function for improved segmentation accuracy.

Summary:

• The study proposes Dual-SAUNet++, a dual-path network incorporating LGDL for segmenting COVID-19 lesions in CT scans.
• LGDL combines generalized Dice loss and mean square error for adaptive weighting.
• The model demonstrates strong performance with high Dice (87.81%) and IoU (79.20%) scores, and excellent specificity (99.83%).

Impact:

• Dual-SAUNet++ offers robust anti-noise capabilities and segments multi-scale lesions effectively.
• Accurate lesion segmentation assists clinicians in assessing COVID-19 severity and planning treatment.
• Provides a reliable basis for clinical decision-making in COVID-19 management.