Abstract

Info

Title:: Automatic segmentation of organs at risk in radiotherapy using deep deconvolutional neural network

Author(s):: MEN Kuo; DAI Jianrong; National Cancer Center/Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing 100021, China

Keywords:: ; deep learning; automatic segmentation; radiotherapy; organs-at-risk segmentation; deep deconvolutional neural network

Abstract:: Objective The segmentation of organs-at-risk (OAR) is essential to radiotherapy. However, the current manual segmentation is time-consuming and dependent on the physicians’ knowledge and experience. Herein deep deconvolutional neural network (DDNN) is proposed for the automatic and accurate segmentation of OAR. Methods DDNN is an end-to-end framework for automatic segmentation. The data of 230 head and neck patients were selected in this study. Among all the selected patients, 184 cases were randomly chosen as training set used for adjusting the parameters of automatic segmentation model, and the other 46 cases were regarded as test set used for evaluating the performance of the proposed method. The OAR to be segmented included brainstem, spinal cord, left parotid, right parotid, left temporal lobe, right temporal lobe, thyroid gland, larynx and trachea. Dice similarity coefficient and Hausdorff distance were used to measure the accuracy of segmentation. Results All the Dice similarity coefficient values of OAR were higher than 0.70, with a mean value of 0.81 and the Hausdorff distance values were within 5.0 mm, with a mean value of 4.2 mm, which demonstrated that the proposed method could segment OAR accurately. Conclusion The DDNN-based automatic method for OAR segmentation achieves an accuracy segmentation result, providing technical supports for the automation of radiotherapy.