MS-UNet: A Novel Multi-scale U-shaped Network for COVID-19 CT Image Segmentation

Abstract

The U-Net network has its own powerful capabilities in medical image segmentation tasks, yet still it is a challenging task to make U-Net accurately segment the infected lesions of COVID-19 CT images because these lesion areas are usually irregular in shape, various in size, and blurry in boundaries. In this paper, a novel multiscale U-shaped network based on U-Net for accurate segmentation of lesion regions in COVID-19 CT images is proposed. First, we generate two auxiliary scale features (f_i^0.5, f_i^1.5) based on the main scale feature (f_i^1.0) through zoom strategy. Secondly, we design the Scale Integration Module (SIM), which is capable of filtering and aggregating scale-specific features and can fully exploit multi-scale semantic information. Thirdly, the hierarchical mixed module (HMM) has successfully substituted for the down-up aggregation process of the U-Net network, which further enhances the mixed scale features. On the dataset COVID-19-CT829, compared with the recent COVID-19 segmentation model, hiformer, the Dice, Sen and F-measure of our network have increased by 2.24%, 2.83%, 3.14%, respectively; on the dataset COVID-19-CT100, the Dice, Sen and F-measure of our network have increased by 2.91%, 3.72%, 2.42%, respectively. Moreover, we have validated the generalizability and portability of our network on other medical datasets (Polyp segmentation dataset: CVC-612 and kvasir), and our network has also achieved superior results of COVID-19 CT image segmentation.