Automatic Segmentation Of Radiotherapy Target For Nasopharyngeal Carcinoma Based On Cascaded Convolutional Neural Network

At present,the segmentation of the target region of radiotherapy for nasopharyngeal cancer still needs to be completed manually by physicians,which is time-consuming and laborious,and the mapping consistency among different physicians is poor.In recent years,the rapid development of automatic segmentation technology based on deep learning can provide a reference for doctors to outline,so as to improve the consistency rate of outline,but the effect of automatic segmentation varies with the network model used.In this study,a cascaded convolutional neural network model based on coarse location and fine segmentation was designed by combining Resnet and Densenet to improve the two-dimensional Unet network.The model was trained and tested with CT images of 224 patients with primary nasopharyngeal carcinoma.In order to improve the training efficiency of the model,the CT images were preprocessed by gray truncation and zero-mean normalization,and the sample number was amplified by the image enhancement technology,and the cross-entropy loss function was added on the basis of the dice loss function.The accuracy of segmentation results was evaluated by Jaccard similarity coefficient(JSC),Dice similarity coefficient(DSC),Hausdorff distance(HD),95%HD,Meas surface distance(MSD)and other parameters.Then,they were compared with Resunet,Attention Unet,MAD-Unet model and ABAS(Atlas based segmentation methods)segmentation results.The results showed that:(1)For GTV target region,DSC and JSC had the highest similarity coefficient,and HD values were worse than those of Resunet model,and 95%HD and MSD were worse than those of Attention Unet model.For the CTV target region,the coefficient indexes of the cascaded convolutional neural network are the best.(2)Except for the HD index,the indexes of ABAS segmentation method were all worse than those of the cascading automatic segmentation model.(3)Cascade automatic segmentation model and the manual sketch of gray histogram feature,gray level co-occurrence matrix and neighborhood grayscale difference matrix have significant correlation(|r| > 0.4),with statistical significance(p < 0.05).The research indicated that the cascaded convolutional neural network model based on deep learning can segment the radiotherapy target area of nasopharyngeal cancer well,and has a higher segmentation accuracy for the CTV target region.Most of the radiomics features of the target region automatically segmented by the cascade convolutional neural network and the target region manually delineated were significantly correlated.