Songcai Yan, Xinjun Hu, Shanshan Hu, Jianping Tian, Qinyuan Xue, Lin He, Jianheng Peng
Due to the relatively small size and complex internal structure of the pancreas, the segmentation is often inaccurate during image processing. A more effective automatic segmentation method is proposed to solve this problem. In order to make full use of encode-decode structure to extract image features, this study cascaded the improved Unet and RSU structure networks to form a new U-shaped network, and carried out feature fusion at the output end of each network. In order to further extract the key semantic information of image features, a hybrid attention mechanism is introduced to enhance the attention of the network to the important regions in the image. In order to reduce network computation and speed up model convergence, all ordinary convolution in cascade network is replaced by deep separable convolution. Compared with the benchmark model Unet, Dice, IoU, Precision, Recall, Specificity and Accuracy were increased by 5.32%, 7.44%, 5.31%, 5.69%, 0.06% and 0.13%, respectively, when trained and tested on NIH data set. This method enhances the ability to perceive and locate pancreatic features, and can effectively improve the accuracy of pancreatic segmentation.