Kuo Men1, Xinyuan Chen1, Ji Zhu1, Bining Yang1, Ye Zhang1, Junlin Yi1, And Jianrong Dai2. 1. National Cancer Center/National Clinical Research Center for Cancer/Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing 100021, China. 2. National Cancer Center/National Clinical Research Center for Cancer/Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing 100021, China. Electronic address: dai_jianrong@cicams.ac.cn.
Abstract
PURPOSE: Convolutional neural networks (CNNs) offer a promising approach to automated segmentation. However, labeling contours on a large scale is laborious. Here we propose a method to improve segmentation continually with less labeling effort. METHODS: The cohort included 600 patients with nasopharyngeal carcinoma. The proposed method was comprised of four steps. First, an initial CNN model was trained from scratch to perform segmentation of the clinical target volume. Second, a binary classifier was trained using a secondary CNN to identify samples for which the initial model gave a dice similarity coefficient (DSC) < 0.85. Third, the classifier was used to select such samples from the new coming data. Forth, the final model was fine-tuned from the initial model, using only selected samples. RESULTS: The classifier can detect poor segmentation of the model with an accuracy of 92%. The proposed segmentation method improved the DSC from 0.82 to 0.86 while reducing the labeling effort by 45%. CONCLUSIONS: The proposed method reduces the amount of labeled training data and improves segmentation by continually acquiring, fine-tuning, and transferring knowledge over long time spans.
PURPOSE: Convolutional neural networks (CNNs) offer a promising approach to automated segmentation. However, labeling contours on a large scale is laborious. Here we propose a method to improve segmentation continually with less labeling effort. METHODS: The cohort included 600 patients with nasopharyngeal carcinoma. The proposed method was comprised of four steps. First, an initial CNN model was trained from scratch to perform segmentation of the clinical target volume. Second, a binary classifier was trained using a secondary CNN to identify samples for which the initial model gave a dice similarity coefficient (DSC) < 0.85. Third, the classifier was used to select such samples from the new coming data. Forth, the final model was fine-tuned from the initial model, using only selected samples. RESULTS: The classifier can detect poor segmentation of the model with an accuracy of 92%. The proposed segmentation method improved the DSC from 0.82 to 0.86 while reducing the labeling effort by 45%. CONCLUSIONS: The proposed method reduces the amount of labeled training data and improves segmentation by continually acquiring, fine-tuning, and transferring knowledge over long time spans.
Authors: Wai Tong Ng; Barton But; Horace C W Choi; Remco de Bree; Anne W M Lee; Victor H F Lee; Fernando López; Antti A Mäkitie; Juan P Rodrigo; Nabil F Saba; Raymond K Y Tsang; Alfio Ferlito Journal: Cancer Manag Res Date: 2022-01-26 Impact factor: 3.989