Pancreas Segmentation Based on Multi-scale Convolution and Attention Module

Abstract

The pancreas is characterized by its small size and variable shape. The current neural network model is not high in segmentation accuracy of pancreatic target region image, and there is still much room for improvement. U-Net is a kind of network with excellent performance in medical image segmentation. However, the traditional U-Net network is not strong enough to extract feature information of small regions. To improve the ability of the traditional U-Net network to extract feature information of local small regions, and to make the pancreas organ segmentation boundary clearer, this paper designs a multi-scale convolution block (MCB), Channel Attention module and Attention Gates structure are added into the U-Net network. The network model emphasizes the information extraction of a small region, so it can improve the segmentation effect. Dice coefficient, MIoU coefficient, and Precision coefficient are selected as the evaluation indexes of the training effect and conducted a series of experiments on the NIH pancreas segmentation dataset. The experimental data show that the model in this paper has improved by 1-2 percentage points in each positive index coefficient compared with the latest model, and the model converged faster, which confirms the superiority of the model.