Pancreatic Tumor Segmentation Based on 3D U-Net with Densely Connected Atrous Spatial Pyramid Module and Attention Module
Jiakun Deng, Yi Mou · 2023
The segmentation of CT images of pancreatic tumors is of great significance for the diagnosis and treatment of pancreatic tumors. Due to the small size and irregular shape of pancreatic tumors, precise segmentation of pancreatic tumor CT images using neural networks remains a challenge in machine learning. At present, the dice coefficient, MIoU coefficient, and Precision coefficient of neural network for pancreatic tumor CT image segmentation are only 50%, 30%, and 50% at most, respectively. There is a lot of room to improve the segmentation effect. This paper proposes a CT image segmentation method for pancreatic tumors based on a 3D U-Net network and integrating Densely Connected Atrous Spatial Pyramid Module (DenseASP Module) and Attention Module (AM). In order to effectively extract small target feature information, this network adds DenseASP to each layer of the downsampling section; At the same time, in order to filter irrelevant information and enhance the extraction of target area information, we add Attention Gates (AG) to each layer of the upsampling section, and Channel Attention Module (CAM) to the skip connection section. The Attention Gates and Channel Attention Modules together constitute the Attention Module (AM) of our network. This paper selects dice coefficient, MIoU coefficient, Precision coefficient as the evaluation index of segmentation effect, and uses the pancreas dataset MSD Pancreas to evaluate our model. The experimental results show that the various segmentation evaluation indicators of the model in this paper have improved by about 2 percentage points compared to the current neural network.