Densely connected deep U-Net for abdominal multi-organ segmentation

The U-Net is the well-known architecture for semantic segmentation and has achieved remarkable successes in many medical image segmentation applications. However, the features learned by standard convolution layers are not distinctive. To address this problem, we propose a novel architecture, called densely connected deep U-Net(DC U-Net). Specifically, the HU values in CT slices were first windowed in a range to exclude irrelevant organs, and then put it into DC U-Net for training. The DC U-Net consists of three blocks with dense connections, selective deconvlution layers with upsample and transconvlution filling function. To further improve the accuracy of small regin of interest segmentation with limited dataset, we proposed a novel loss function. With respect to the ground truth, average Dice overlap ratios for the liver and spleen are 94.9% and 92.1% respectively. The results demonstrated its potential in clinical usage with high effectiveness, robustness and efficiency.