| Zhang Shuming,Yang Ruijie,Zhu Senhua,Wang Hao,Tian Suqing,Zhang Xuyang,Li Jiaqi,Lei Runhong.Comparative study of two different methods for automatic segmentation of organs at risk in head and neck region[J].Chinese Journal of Radiological Medicine and Protection,2020,40(5):385-391 |
| Comparative study of two different methods for automatic segmentation of organs at risk in head and neck region |
| Received:December 23, 2019 |
| DOI：10.3760/cma.j.issn.0254-5098.2020.05.010 |
| KeyWords:Automatic segmentation Organs at risk Deep learning Atlas library |
| FundProject:国家自然科学基金（81071237，81372420） |
| Author Name | Affiliation | E-mail | | Zhang Shuming | Department of Radiation Oncology, Peking University Third Hospital, Beijing 100191, China | | | Yang Ruijie | Department of Radiation Oncology, Peking University Third Hospital, Beijing 100191, China | ruijyang@yahoo.com | | Zhu Senhua | Beijing Linking Medical Technology Co., Ltd, Beijing 100085, China | | | Wang Hao | Department of Radiation Oncology, Peking University Third Hospital, Beijing 100191, China | | | Tian Suqing | Department of Radiation Oncology, Peking University Third Hospital, Beijing 100191, China | | | Zhang Xuyang | Department of Radiation Oncology, Peking University Third Hospital, Beijing 100191, China | | | Li Jiaqi | Department of Radiation Oncology, Peking University Third Hospital, Beijing 100191, China | | | Lei Runhong | Department of Radiation Oncology, Peking University Third Hospital, Beijing 100191, China | |
|
| Hits: 2202 |
| Download times: 1108 |
| Abstract:: |
| Objective To develope a deep-learning-based auto-segmentation model to segment organs at risk (OARs) in head and neck (H&N) region and compare with atlas-based auto-segmentation software (Smart segmentation). Methods The auto-segmentation model consisted of classification model and segmentation model based on deep learning neural network. The classification model was utilized to classify CT slices into six categories in the cranio-caudal direction, and then the CT slices corresponding to the categories for different OARs were pushed to the segmentation model respectively. The CT image data of 150 patients were used for auto-segmentation model training and building atlas library in Smart segmentation software. Another 20 patients were used as testing dataset for both auto-segmentation model and Smart segmentation software. Dice similarity coefficient (DSC) and Hausdorff distance (HD) were used to evaluate the accuracy of two method, and auto-segmentation time cost was recorded. Paired Student's t-test or non-parametric Wilcoxon signed-rank test was performed depending on result of normality test. Results The DSC and HD of auto-segmentation model for brainstem, left eye, right eye, left optic nerve, right optic nerve, left temporal lobe, right temporal lobe, mandible, left parotid and right parotid were 0.88 and 4.41 mm, 0.89 and 2.00 mm, 0.89 and 2.12 mm, 0.70 and 3.00 mm, 0.80 and 2.24 mm, 0.81 and 7.98 mm, 0.84 and 8.82 mm, 0.89 and 5.57 mm, 0.70 and 11.92 mm, 0.77 and 11.27 mm respectively. The results of auto-segmentation model were better than those of Smart segmentation (t=3.115-7.915, Z=-1.352 to -3.921, P<0.05) except left and right parotids. In addition, the speed of auto-segmentation model was 51.28% faster than that of Smart segmentation. Conclusions In this study, the deep-learning-based auto-segmentation model demonstrated superior performance in accuracy and efficiency on segmenting OARs in H&N CT images, which was better than Smart segmentation software. |
| HTML View Full Text View/Add Comment Download reader |
| Close |
|
|
|
