Rectal Tumor Segmentation Based On Deep Neural Networks

Rectal cancer is a pre-stage feature of rectal cancer and an important basis for rectal cancer.Therefore,using algorithms to accurately segment the location and size of rectal tumors is a challenging task.In the process of diagnosing rectal cancer,the rectal tumor area is automatically and accurately segmented from the CT image,which helps to improve the speed and accuracy of the doctor's diagnosis.Rectal tumor segmentation is a very challenging technique,and it is also a significant work in medical image segmentation.Therefore,in this paper,aiming at the problem of rectal tumor segmentation,an automatic segmentation algorithm of rectal tumor based on U-Net improved model is proposed.First,embed sub-encoding modules in each level of the encoder of the U-Net model to improve the feature extraction capabilities of the model;secondly,by comparing the optimization performance of different optimizers,the most suitable optimizer is obtained for training the model;finally,the training set is expanded using data.The model is more fully trained,thereby improving the segmentation performance of the model.Ours proposed a comparison experiment between the improved model U-Net-SCB and the three deep neural network models of U-Net,Y-Net and Focus Net Alpha.The segmented area obtained by the improved model proposed in this paper is closer to the real tumor area,and the segmentation performance for small targets is more prominent.The precision rate P,the recall rate R and the Dice coefficient are better than the other three models.Due to the improved U-Net deep neural network algorithm,the accuracy of segmenting rectal tumors needs to be further improved.Therefore,this article designs different deep neural network models to improve the segmentation accuracy of the deep neural network algorithm.The improved model can effectively segment the rectal tumor area,which is an intentional exploration of the automatic segmentation technology of rectal tumor area.In order to improve the ability to extract sufficient feature information in rectal tumor segmentation,this paper proposes a dual-path U-Net network with a covariance self-attention mechanism(CSA-DPUNet).The proposed network mainly includes two improvements: 1)The U-Net with only one path structure is modified to include two contraction paths and two expansion paths(the network model is called DPUNet),so that CT images can be obtained extract more feature information in DPUNet;2)Introduce the cross self-attention module in DPUNet,and use covariance operation to replace the original correlation calculation method to further improve the feature expression ability of DPUNet and improve the accuracy of rectal tumor segmentation.The experimental results show that,compared with U-Net-SCB model,the Dice coefficient,P,R and F1 of CSA-DPUNet are increased by 15.31%,7.2%,11.8% and 9.5%,respectively,which proves that the CSA-DPUNet proposed is effective in segmentation of rectal tumors.The positive effect is of reference significance for smart medical treatment.