Robust COVID-19 Lung and Infection Segmentation in Computed Tomography Scans via U-Net with ResNet-SEResNet Encoders and Optimized Sliding Window Preprocessing

Accurate segmentation of lung and infection regions in CT scans plays a vital role in the early diagnosis and clinical management of COVID-19. This study proposes a robust and efficient DL approach utilizing a U-Net architecture equipped with ResNet and SEResNet encoder backbones, combined with optimized preprocessing strategies. The integration of flipping and sliding window techniques, which substantially increased the data volume and enhanced feature extraction at multiple scales, is a key innovation of this work. Unlike many prior studies that focused solely on infection segmentation, this study addresses both lung and infection region segmentation on the publicly available COVID-19 CT dataset. Among the various backbones tested, SEResNet18 was selected as the optimal choice due to its high accuracy and computational efficiency.The proposed model achieved outstanding segmentation performance, with Dice score, F1-score, and IoU reaching 0.9897, 0.9900, and 0.9796 for lung segmentation and 0.9339, 0.9357, and 0.8770 for infection segmentation, respectively. These results not only surpass those of previous studies using the same dataset but also highlight the significance of TP in improving model generalizability. This study contributes to the field by demonstrating that carefully designed data preparation pipelines can be as impactful as architectural innovations, paving the way for high-performance, resource-efficient segmentation systems applicable to broader medical imaging tasks.