Optimized attention U-Net for enhanced lung and area of infection segmentation in chest X-Rays and CT scans

Medical image segmentation is crucial for accurate diagnosis, effective treatment planning, and thorough disease monitoring. However, segmentation inaccuracies can have serious clinical consequences clinically significant errors occur in up to 18 % of automated hippocampal segmentations and diagnostic errors in radiology including misdiagnoses, missed diagnoses, and delayed diagnoses affect 10–26 % of cases. Deep learning models, particularly the U-Net architecture, have significantly advanced this field, demonstrating flexibility and success across various medical imaging tasks. However, despite enhancements to U-Net's backbone, bottleneck, and skip connections, existing variants often struggle to capture both global context and fine-grained details due to the limitations of convolutional operations. To address these challenges, we propose a novel residual U-Net architecture that combines multiple attention mechanisms to improve feature representation. Specifically, Efficient Channel Attention (ECA) is integrated into the encoder to recalibrate channel-wise feature responses, while the Dual Attention Network (DANet) is applied in the bottleneck to capture long-range spatial and channel dependencies. Attention gates at the skip connections further refine feature selection during up-sampling. We evaluated our proposed architecture on the public COVID-19 CT dataset and the COVID-QU-Ex dataset. Experimental results demonstrate that our U-Net variant, specifically the ECA-Attention with Dual Attention Network (DANet), outperforms several state-of-the-art segmentation methods, including the SE-Attention with DANet model. On unseen data from the COVID-QU-Ex dataset, our model achieved a mean Intersection over Union (IoU) of 0.9740, a mean Dice Coefficient of 0.9784, a loss of 0.0253, and an accuracy of 0.9906 for lung segmentation. For infection area segmentation, the model achieved a mean IoU of 0.8272 and a mean Dice Coefficient of 0.7837. These results indicate a significant improvement in segmentation accuracy on unseen data, highlighting the superior capability of our approach in medical image segmentation and its potential for broader clinical applications.