A modified U-Net method for pneumonia lesion segmentation based on coordinate attention mechanism

The outbreak of COVID-19 has posed significant challenges to the global medical system, underscoring the importance of automatically and accurately segmenting the focal areas of COVID-19 from CT images for effective treatment. This paper addresses the limitations of traditional manual segmentation, which is heavily reliant on the operator’s experience, and the difficulties faced by the original U-Net model in delineating fuzzy boundaries of infection sites in CT scans. We introduce a coordinate attention mechanism to adaptively adjust the network’s focus on different shapes and sizes of the targeted areas, thereby enhancing the accuracy of COVID-19 CT image segmentation. Our results demonstrate that the segmentation method enhanced by the coordinate attention mechanism significantly outperforms the original U-Net, achieving MIoU, Precision, and Recall metrics of 80.99%, 88.27%, and 89.77%, respectively. These figures represent improvements of 2.27%, 3.64%, and 0.79% over the original U-Net model, illustrating the efficacy of our approach in improving the precision of medical imaging analysis in the context of COVID-19.