DTUnet: A Transformer-based UNet Combined with DenseASPP Module for Pancreas Segmentation

Accurate pancreas segmentation is of great significance for the diagnosis and treatment of pancreatic cancer. Exploiting FCN, UNet or their variants, these CNN-based methods, to complete this task has become the de-facto standard and great success has been achieved. However, convolutional operation fails to build long-range dependency which is crucial for segmentation, hindering the further development of CNN-based methods. To address the difficulty, we propose the DTUnet network, which introduces the DenseASPP module and Transformer on the basis of UNet and stacks the two in a sequential manner. Transformer connects each pixel of the input feature maps to generate a global receptive field, thus capturing the global context information and realizing the construction of long-range dependency. Meanwhile, to alleviate the training challenges of Transformer's data-hungry, DTUnet employs DenseASPP module to generate rich and multi-scale high-level semantic feature maps as the input of Transformer, ensuring that Transformer can fully leverage global modeling capability even when applied to a small size pancreas segmentation dataset. Benefiting from the combination of the two, the proposed DTUnet generates more efficient and reliable global context information, and ultimately achieves an average Dice coefficient score of 84.77% ±4.65 on the public NIH pancreas segmentation dataset, which is 1.87% higher than UNet. The result is also higher than advanced methods in recent years, indicating that DTU net has the potential to assist doctors to segment the pancreas in clinical application.