Automatic Segmentation Of Organs-at-risks In Head And Neck CT Images Based On Deep Learning

This study aims to develop an automatic segmentation framework with high precision for OARs in head and neck(H&N)CT images using the U-Net neural network of deep learning,so as to realize the automatic segmentation of H&N OARs.The CT images of 200 H&N cancer patients treated with radiotherapy were selected in Fujian Provincial Cancer Hospital(FPCH).The OARs(Body,brainstem,spinal cord,left and right eyeballs,left and right lens,left and right mandibles,left and right temporal lobes)in each set of CT images were contoured by one experienced radiation oncologist according to the contouring guidance from the FPCH.The contour labels of OARs and corresponding original CT images in each case were extracted and saved as data sets in image format.Firstly,the image data in the data sets were preprocessed to meet the training requirements of neural network,and 180 cases were randomly selected as the training data sets,and the remaining 20 cases were selected as the test data sets.Then,an automatic segmentation neural network model based on U-Net was constructed,and the training sets(including contour labels and corresponding CT images)and test sets(including CT images only)of each OARs were input into the neural network one by one for training and debugging.The predicted results of automatic CT segmentation for the test sets were output and the training model was saved.A series of post-processing were performed to obtain the contours of the OARs in the test sets.Finally,the Dice coefficient and Hausdorff distance of the predicted contour graph and the label contour graph were calculated,and the data of this study were compared with the data in relevant literatures for evaluating the performance of the automatic segmentation model,and the clinical feasibility of the automatic segmentation method in this study.The results showed that the accuracy of automatic segmentation to the 20 test cases was comparable to manual results.The Dice coefficient of these OARs(Body,brainstem,spinal cord,left eye globe,right eye globe,left lens,right lens,left mandible,right mandible,left temporal lobe,right temporal lobe)were 0.95±0.02,0.86±0.02,0.87±0.03,0.85±0.02,0.86±0.02,0.80±0.03,0.81±0.02,0.88±0.02,0.89±0.01,0.70±0.03,0.71±0.02,respectively.The Hausdorff distance were 2.8±0.9mm,3.9±1.5mm,3.8±1.3mm,4.1±1.2mm,4.2±1.1mm,4.5±1.3mm,4.4±1.5mm,3.7±1.3mm,3.6±1.2mm,5.9±1.8mm,5.5±1.6mm,respectively.Compared with data from literatures,it was found that the Dice coefficient of these OARs were very close to these literatures expect the left and right temporal lobes.Meanwhile,the average time for automatic segmentation of each CT was only 78.27±12.36 milliseconds,and the average time for complete segmentation of a head and neck OARs is 18.6±2.1 seconds,which has a high segmentation efficiency.The study shows that compared with manual delineation of OARs,the automatic segmentation programme of this study can bring higher accuracy and much lower time consumption and less burden of doctors' work.Therefore,this study shows that the automatic segmentation program based on U-Net neural network is feasible for clinical application.