Click here to


Are you sure ?

Yes, do it No, cancel

Lifelong Learning for Clinical Target Segmentation of Nasopharyngeal Cancer with Fewer Labeling

K Men*, X Chen, Y Zhang, J Zhu, J Yi, J Dai, National Cancer Center/National Clinical Research Center for Cancer/Cancer Hospital, Chinese Academy of Medical Sciences, Beijing, 100021, China,


(Sunday, 7/12/2020)   [Eastern Time (GMT-4)]

Room: AAPM ePoster Library

Purpose: Convolutional neural networks (CNNs) offer a promising approach to automating segmentation. However, labeling large-scale contours is laborious and time-consuming. Moreover, when trained with new data, the models may deteriorate as previously learned knowledge is forgotten. In this study, we proposed a lifelong learning method for segmentation with fewer labeling.

Methods: A total of 600 patients with nasopharyngeal carcinoma were enrolled in this study, divided into a training set (250 cases), a validation set (50 cases), a re-training set (250 cases), and a test set (50 cases) in chronological order. A state-of-the-art CNN was trained to perform segmentation of the clinical target volume. The proposed lifelong learning comprised four steps. First, we trained Model_A from scratch with the training set. Second, we trained a binary classifier using a secondary CNN with the validation set. It was designed to identify samples on which Model_A had a Dice similarity coefficient (DSC) below 0.85. Third, the classifier was used to select such samples from the re-training set. Finally, transfer learning was then used to train Model_B by fine-tuning Model_A with only the selected samples from the re-training set. A Model_C was fine-tuned from Model_A with all the samples from the re-training set for comparison.

Results: The classifier can detect the poor segmentation of the model with an accuracy of 91%. When Model_A, Model_B, and Model_C were applied to the test set, the DSCs were 0.83 ± 0.03, 0.87 ± 0.02, and 0.87 ± 0.02, respectively. With the classifier, labeling 56% slices were enough to achieve comparable results; Lifelong learning improved the segmentation accuracy from 0.83 (Model_A) to 0.87 (Model_B) and reduced training time by up to 33%.

Conclusion: The proposed method can reduce the amount of labeled training data, improve segmentation by continually acquiring, fine-tuning, and transferring knowledge over long time spans.

Funding Support, Disclosures, and Conflict of Interest: This work was supported by the Beijing Hope Run Special Fund of Cancer Foundation of China (LC2019B06, LC2018A14), the Beijing Municipal Science & Technology Commission (Z181100001918002), and the National Natural Science Foundation of China (11975313).




IM/TH- Image Segmentation Techniques: General (Most aspects)

Contact Email