Research On Automatic Segmentation And Accuracy Of Heart Substructure Based On Convolutional Neural Network

Objective:Using deep convolutional neural networks with different loss functions to segment cardiac substructures on CT enhanced positioning for radiotherapy,and compare the segmentation accuracy,and contribute to the application of automatic segmentation of cardiac substructures in future clinical work.Methods:We collected data of 47 patients with thoracic tumors admitted to the Department of Radiotherapy of Yunnan Cancer Hospital from May 1,2020 to June 30,2020.Each patient has a set of location-enhanced CT images,in which organs at risk are defined as 10 structures including the heart and its substructures(pericardium,heart,left atrium,left ventricle,right atrium,right ventricle,left main stem,left anterior Descending artery,left circumflex artery,right coronary artery).First,a radiation oncologist who has been engaged in radiation oncology for more than two years,based on experience and guidelines,outlines ten endangered organs on the enhanced CT,and uses the endangered organs manually drawn by the radiation oncologist as the standard.The 47 patients were randomly divided into training set,validation set,and test set.Among them,29 cases were training set,3 cases were validation set,and 15 cases were test set.Then use four deep convolutional neural networks(GDL U-Net,WCEGDL U-Net,ELL U-Net,GDL V-Net)that combine different loss functions for training,and use the trained model to locate the CT The organ is automatically segmented,and finally the results of the automatic segmentation are compared with the content manually outlined by the radiation oncologist.Using dice similarity coefficient(DSC),Hausdorff distance(HD),Jaccard coefficient(JC),volume difference(VD)as quantitative evaluation indicators,the four automatic segmentation results are compared with manually drawn reference valuesResults:The DCNN network that combines different loss functions has achieved a segmentation DSC of more than 0.87 for the pericardium,heart,and four chambers,and the VD value is all below 0.1,the volume difference is not large,and the segmentation result is ideal.The JC values of pericardium and heart segmentation were all above 0.91,and the contour similarity was good.Among them,the segmented DSC of the pericardium by WCEGDL U-Net reaches 0.963,95%HD is 3.447mm;ELL U-Net segmentation DSC of the heart reaches 0.967,95%HD is 3.476mm;GDL U-Net segmentation of left atrium and right ventricle is better,DSC is 0.897(95%HD:3.423)mm),0.914(95%HD:4.240mm);GDL V-Net has better segmentation performance for the right atrium and left ventricle,with a DSC of 0.882(95%HD:3.902mm),0.942(95%HD:2.820mm)).The four DNNs have their own advantages in segmenting different substructures of the heart.However,in the segmentation of small organs,the DSC of the four DCNNs did not reach above 0.7,and only the segmentation for LAD was closer to 0.7.Conclusion:The GDL U-Net,WCEGDL U-Net,ELL U-Net,and GDL V-Net used in this study have achieved better segmentation effects on the pericardium,heart and four chambers in the cardiac substructure segmentation,but in the coronary artery segmentation.The segmentation effect is still not ideal,and the deep convolutional neural network model needs to be further improved to enhance the segmentation performance of small organs.