A Comparative Study on Image Segmentation Models in COVID-19 Diagnosis

With Omicron sweeping the world, it has brought huge pressure on the healthcare system, and quick diagnosis of pneumonia caused by COVID-19 using chest computed tomography (CT) scans plays a key role in saving lives. Image processing techniques have been widely used to analyze CT scans and other medical images, which also facilitate the diagnosis and treatment of Coronavirus disease. This article introduces several image segmentation models that are used to facilitate the diagnosis of COVID-19 and applies several recently proposed deep learning (DL) models to the task, including UNEt TRansformers (UNETR++) and Dual Attention-guided Efficient Transformer (DAE-Former). We compare the performance of these models and provide a thorough analysis of different methods. The experimental results show that for the task of segmenting COVID-19 lesion areas, both Transformer-based models obtained better performance in terms of the value of mIoU compared with U-Net, a CNN-based model, while U-Net obtained higher accuracy. The DAE-Former model has superior anti-noise ability to UNETR++, whereas UNETR++ is more robust in terms of domain transfer.