A Novel Full-Scale Skip Connections Approach Based on U-Net for COVID-19 Lesion Segmentation in CT Images

In the post-pandemic era, as COVID-19 continues to spread, CT imaging is indispensable for diagnosing COVID-19. Utilizing computer vision techniques to segment the lesion regions in CT scans can assist doctors in efficient and accurate diagnosis. However, traditional CNN-based U-net segmentation models are more adept at extracting local information, lacking overall awareness of the data, and suffering from semantic loss in the upsampling and downsampling process. To tackle these concerns, we present a Transformer-based full-scale skip connections Unet model. By transforming the traditional CNN structure into a SwinTransformer structure, the model can focus more on the global information of the image, making the instance features more robust and informative. Additionally, we incorporate full-scale skip connections to facilitate the upsampling module to simultaneously access the spatial information from each downsampling module, reducing spatial information loss and improving the segmentation accuracy of the model. We trained and tested our model using an independent dataset of COVID-19 from Wuhan. Experimental results demonstrate that our model exhibits good segmentation capability for COVID-19 lesions and outperforms other methods in terms of average precision. Furthermore, we performed ablation experiments for validation. The effectiveness of the full-scale skip connections.