CoviSwin: A Deep Vision Transformer for Automatic Segmentation of COVID-19 CT Scans

Precise segmentation of COVID-19 lesions in chest Computed Tomography (CT) scans can directly impact patient care, yet existing methods struggle, when undertaking this task, with the heterogeneous appearance of ground-glass opacities, consolidations, and the availability of limited labeled data. We propose herein CoviSwin, a Transformer-based U-shaped encoder-decoder network that combines the Large model of Swin Transformer Version 2 with attention and residual connections to capture both global context and fine details. A two-phase training strategy is applied whereby in the first phase the encoder is initially frozen while training the decoder on the public SemiSeg dataset, then in the second phase, the encoder is partially unfrozen while the whole model is trained on the publicly available MedSeg dataset. The model achieves a ten-run mean sensitivity value of 0.790 ± 0.012, an average Dice Similarity Coefficient (DSC) score of 0.781 ± 0.0068, and an average specificity of 0.962 ± 0.0049, outperforming the sensitivity results obtained by recent models such as NextSeg of 2024 and GFNet of 2022 by 8.07% and 7.48%, respectively. These findings demonstrate the potential of CoviSwin as an effective model for clinical COVID-19 lesion segmentation.