Evaluation of Deep Learning Approaches for Segmentation of COVID-19 Lung CT Scans

The COVID-19 pandemic, initiated by the emer-gence of the SARS-Co V-2 virus in late 2019, rapidly be-came a global crisis, straining healthcare systems worldwide. While laboratory tests are primary for COVID-19 diagnosis, computed tomography (CT) scans are valuable for detecting lung abnormalities associated with the virus such as ground-glass opacity and consolidation. However, manual annotation by radiologists is time-consuming and error-prone, necessitating automated segmentation. This study compares deep learning (DL) models for automatically segmenting COVID-19-related lung CT images. The research assesses four image segmentation models using important metrics such as Accuracy, Sensitivity, Specificity, Intersection over Union, and Dice coefficient. This finding of the combination of MobileNetV3 and U-Net to be the most effective overall with an accuracy of 0.97 and a Dice of 0.96. The application of Contrast limited adaptive histogram equalization (CLAHE) consistently improves segmentation performance across models, underscoring the need for further research to refine segmentation techniques for diverse datasets and tasks.