RTUNet: Residual transformer UNet specifically for pancreas segmentation

Abstract

Accurate pancreas segmentation is crucial for the diagnostic assessment of pancreatic cancer. However, large position changes, high variability in shape and size, and the extremely blurred boundary make the task of pancreas segmentation challenging. To alleviate these challenges, we propose the residual transformer UNet (RTUNet) to fit the nature of the pancreas. Specifically, a residual transformer block is implemented to extract multi-scale features from a global perspective which captures high variabilities in the pancreas position. In addition, a dual convolutional down-sampling strategy is leveraged to obtain precise shape and size features of the pancreas in a large receptive field which prevents the loss of information. We finally propose a dice hausdorff distance loss that makes the network focus on the pancreas boundary. Through extensive experiments on the public NIH dataset, we achieved a dice similarity coefficient (DSC) of 86.25%, which outperforms the state-of-the-art DSC of 85.49%. In addition, our method surpasses the baselines by more than 3.0% on DSC and improves the min DSC by 2.93%. Furthermore, ablation studies are also performed to prove the effectiveness of each proposed module.