A comparison of U-Net series for CT pancreas segmentation

The problems of the large variation in shape and location, and the complex background of many neighboring tissues in the pancreas segmentation hinder the early detection and diagnosis of pancreatic diseases. The U-Net family achieve great success in various medical image processing tasks such as segmentation and classification. This work aims to comparatively evaluate 2D U-Net, 2D U-Net++ and 2D U-Net3+ for CT pancreas segmentation. More interestingly, We also modify U-Net series in accordance with depth wise separable convolution (DWC) that replaces standard convolution. Without DWC, U-Net3+ works better than the other two networks and achieves an average dice similarity coefficient of 0.7555. Specifically, according to this study, we find that U-Net plus a simple module of DWC certainly works better than U-Net++ using redesigned dense skip connections and U-Net3+ using full-scale skip connections and deep supervision and can obtain an average dice similarity coefficient of 0.7613. More interestingly, the U-Net series plus DWC can significantly reduce the amount of training parameters from (39.4M, 47.2M, 27.0M) to (14.3M, 18.4M, 3.15M), respectively. At the same time, they also improve the dice similarity compared to using normal convolution.