Study On The Method Of Automatically Describing The Clinical Target Volume Of Rectal Cancer CT Images Based On U2-Net Network

Objective: Segmentation of rectal cancer by Computed Tomography(CT)images is an important step in radiotherapy.The shape changes of rectal cancer are different,and the difference between the shape of some nearby organs and rectal cancer is small.These objective factors obviously affect the accuracy of rectal cancer segmentation,leading to the difference of segmentation.Therefore,the description of Clinical Target Volumes(CTV)and Organs At risk(OARs)is very important in clinical division,while manual drawing of CTV and Organs At Risks(OARs)is very time-consuming.In addition,because there are borderless cancer cells,tumor spread or subclinical disease tissues in CTV of rectal cancer [1],traditional methods are not effective in depicting CTV of rectal cancer.We propose a multi-attention-based U2-Net network to provide a clearer profile for colorectal cancer target segmentation.Methods: For the CT images of 70 rectal cancer patients collected,we used the U2-Net deep network architecture based on multiple attention to perform semantic segmentation.This network structure is a double-layer nested U-shaped network,and multiple attention mechanisms,global pooling layer and edge detection module are added to improve the accuracy and performance of the model.In order to evaluate the characterization performance of the model,we used DSC(similarity coefficient of dice),Mean IOU(mean crossover ratio),F1-Score(harmonic average of accuracy rate and recall rate),MAE(mean loss coefficient)and other evaluation indexes as evaluation criteria for the characterization of CTV and the segmentation of organs at risk.Results: The evaluation results of our method were: DSC = 0.947,Mean IOU = 0.945,F1-Score = 0.951,MAE = 0.007.In addition,compared with the U2-Net network before improvement,the edge rendering performance of the improved network CTV has been significantly improved,and the overall rendering performance of the model has also been improved.In addition,compared with the existing CTV characterization methods: ResU-Net,Deeplabv3+,DDUnet,this study has obvious advantages in the characterization performance of CTV.Conclusions: This paper presents a semantic segmentation method named U2-Net,which uses a multiple attention mechanism to improve the accuracy and efficiency of the target volume(CTV)characterization of rectal cancer,while significantly reducing the number of parameters in the model.At the same time,the introduction of global pooling layer and edge prediction module can improve the edge description performance of the model,and finally achieve more accurate overall description results.