OBJECTIVE: To establish an algorithm based on 3D convolution neural network to segment the organs at risk (OARs) in the head and neck on CT images. METHODS: We propose an automatic segmentation algorithm of head and neck OARs based on V-Net. To enhance the feature expression ability of the 3D neural network, we combined the squeeze and exception (SE) module with the residual convolution module in V-Net to increase the weight of the features that has greater contributions to the segmentation task. Using a multi-scale strategy, we completed organ segmentation using two cascade models for location and fine segmentation, and the input image was resampled to different resolutions during preprocessing to allow the two models to focus on the extraction of global location information and local detail features respectively. RESULTS: Our experiments on segmentation of 22 OARs in the head and neck indicated that compared with the existing methods, the proposed method achieved better segmentation accuracy and efficiency, and the average segmentation accuracy was improved by 9%. At the same time, the average test time was reduced from 33.82 s to 2.79 s. CONCLUSIONS The 3D convolution neural network based on multi-scale strategy can effectively and efficiently improve the accuracy of organ segmentation and can be potentially used in clinical setting for segmentation of other organs to improve the efficiency of clinical treatment.
Head ; Humans ; Image Processing, Computer-Assisted ; Neck ; Neural Networks, Computer ; Organs at Risk ; Tomography, X-Ray Computed

Objective:To validate the feasibility of a deep learning-based clinical target volume (CTV) auto-segmentation algorithm for cervical cancer in clinical settings.Methods:CT data sets from 535 cervical cancer patients were collected. CTVs were delineated according to RTOG and JCOG guidelines, reviewed by experts, and then used as reference contours for training (definitive 177, post-operative 302) and test (definitive 23, post-operative 33). Four definitive and 6 post-operative cases were randomly selected from the testing cohort to be manually delineated by junior, intermediate, senior doctors, respectively. Dice coefficient (DSC), mean surface distance (MSD) and Hausdorff distance (HD) were used for test and comparison between auto-segmentation and RO delineation. Meantime, auto-segmentation time and manual delineation time were recorded.Results:Auto-segmentation models of dCTV 1, dCTV 2 and pCTV 1 were trained with VB-Net and showed good agreement with reference contours in the testing cohorts (DSC, 0.88, 0.70, 0.86 mm; MSD, 1.32, 2.42, 1.15 mm; HD, 21.6, 22.4, 20.8 mm). For dCTV 1, the difference between auto-segmentation and all three groups of doctors was not significant ( P>0.05). For dCTV 2 and pCTV 1, auto-segmentation was better than the junior and intermediate doctors (both P<0.05). Auto-segmentation time consumption was considerably shorter than that of manual delineation. Conclusions:Deep learning-based CTV auto-segmentation algorithm for cervical cancer achieves comparable accuracy to manual delineation of senior doctors. Clinical application of the algorithm can contribute to shortening doctors′ manual delineation time and improving clinical efficiency. Furthermore, it may serve as a guide for junior doctors to improve the consistency and accuracy of cervical cancer CTV delineation in clinical practice.