Multiple Organ Segmentation Based On U-Net Combined Attention Mechanisms

In contemporary society,lung cancer and esophageal cancer have become one of the diseases with the highest morbidity and mortality.Radiation therapy has become one of the main medical treatments for cancer.In the process of radiotherapy,the accurate delineation of the target organs area and the organs at risk in the chest is crucial for physicists to formulate radiotherapy plans.The traditional manual delineation method not only has a huge workload,and wastes the doctor’s time,but also because of the difference between each doctor’s knowledge system,makes it difficult to delineate with a unified standard,which shows that accurate automatic segmentation for medical images is particularly important.Therefore,it is very important to present fast accurate segmentation method.Due to its deep convolutional structure,U-Net network has better segmentation effect on organs with relatively large area,and the segmentation effect of organs with smaller area is poor.For the multi-organ segmentation task of the chest,due to the different sizes,positions and shapes of the chest organs,it is difficult to extract the feature information of multiple organs.In the existing research,there are few studies on the use of other attention mechanisms for chest organ segmentation..Therefore,this paper improves the U-Net network by combining the attention mechanism to achieve accurate segmentation of multiple organs in the chest and improve the segmentation accuracy of small organs.In this paper,the existing Att U-Net network is firstly studied for chest multi-organ segmentation.The Att U-Net network,as a combination of the Attention Gate(AG)and the U-Net network,is rarely used for chest multi-organ segmentation.At the same time,due to the different sizes of the main organs of the chest,when using U-Net for deep feature extraction,the performance of the features of each organ in different layers will be different,which will bring redundant underlying features and disrupt to a certain extent.feature information of small organs,thereby reducing the segmentation accuracy of small organs.In order to improve the segmentation effect of multiple organs,this paper conducts an experimental study on the skip connections of the attention mechanism in Att U-Net to determine the effectiveness of skip connections.Since Att U-Net achieves feature extraction for key target pixels by fusing the feature information of different channels during the implementation process,it does not consider the spatial dependencies between pixels,which has certain limitations in improving the accuracy of organ segmentation.The cross attention mechanism CC(Criss Cross Attention)can obtain the context information of the whole image from all pixels through the loop operation,and generate new features with dense and rich context information.Pay attention to performance.In this paper,the CC module is introduced into the U-Net network,and a new network CU-Net network is proposed to achieve the segmentation of multiple organs in the chest.Further,in order to better utilize the contextual relationship of pixels in the image and focus on key target pixels at the same time,this paper proposes an attention module that combines cross-attention and attention gate to obtain important pixels in the image that consider spatial dependencies.spatial information for more efficient feature extraction.Combining it with U-Net,a novel and efficient CAU-Net network is proposed.This paper uses the public chest multi-organ segmentation challenge seg THOR dataset to conduct experimental research on different networks and different attention combinations.The research results show that,compared with the chest multi-organ segmentation methods proposed by other researchers,CAU-Net is superior to the segmentation effect of other 2D networks,and can reach or slightly better than the segmentation effect of 3D networks,and the average segmentation accuracy can reach 0.9299.Among the several networks based on U-Net that combine different attention mechanisms,the segmentation effect of the CU-Net network is better than that of the Att U-Net network,while the CAU-Net combining CC and AG can get the best segmentation effect..These results show that the network proposed in this paper can improve the average segmentation accuracy in the multi-organ segmentation of the chest,especially improve the segmentation accuracy of small organs,and achieve automatic,fast and accurate segmentation.