Premium
Boosting‐based cascaded convolutional neural networks for the segmentation of CT organs‐at‐risk in nasopharyngeal carcinoma
Author(s) -
Zhong Tao,
Huang Xia,
Tang Fan,
Liang Shujun,
Deng Xiaogang,
Zhang Yu
Publication year - 2019
Publication title -
medical physics
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.473
H-Index - 180
eISSN - 2473-4209
pISSN - 0094-2405
DOI - 10.1002/mp.13825
Subject(s) - segmentation , artificial intelligence , convolutional neural network , nasopharyngeal carcinoma , computer science , pixel , boosting (machine learning) , pattern recognition (psychology) , test set , sørensen–dice coefficient , percentile , radiation therapy , deep learning , artificial neural network , image segmentation , medicine , radiology , mathematics , statistics
Purpose Accurately segmenting organs‐at‐risk (OARs) is a key step in the effective planning of radiation therapy for nasopharyngeal carcinoma (NPC) treatment. In OAR segmentation of the head and neck computed tomography (CT), the low‐contrast and surrounding adhesion tissues of the parotids, thyroids, and optic nerves result in the difficulty in segmentation and lower accuracy of automatic segmentation for these organs than the other organs. In this paper, we propose a cascaded network structure to delineate these three OARs for NPC radiotherapy by combining deep learning and Boosting algorithm. Materials and methods The CT images of 140 NPC patients treated with radiotherapy were collected, and each of the three OAR annotations was respectively delineated by an experienced rater and reviewed by a professional radiologist (with 10 yr of experience). The datasets (140 patients) were divided into a training set (100 patients), a validation set (20 patients), and a test set (20 patients). From the Boosting method for combining multiple classifiers, three cascaded CNNs for segmentation were combined. The first network was trained with the traditional approach. The second one was trained on patterns (pixels) filtered by the first net. That is, the second machine recognized a mix of patterns (pixels), 50% of which was accurately identified by the first net. Finally, the third net was trained on the new patterns (pixels) screened jointly by the first and second networks. During the test, the outputs of the three nets were considered to obtain the final output. Dice similarity coefficient (DSC), 95th percentile of the Hausdorff distance (95% HD), and volume overlap error (VOE) were used to assess the method performance. Results The mean DSC (%) values were above 92.26 for the parotids, above 92.29 for the thyroids, and above 89.37 for the optic nerves. The mean 95% HDs (mm) were approximately 3.08 for the parotids, 2.64 for the thyroids, and 2.03 for the optic nerves. The mean VOE (%) values were approximately 14.16 for the parotids, 14.94 for the thyroids, and 19.07 for the optic nerves. Conclusions The proposed cascaded deep learning structure could achieve high performance compared with existing single‐network or other segmentation algorithms.