AUTHOR=Chen Xinyuan , Zhu Ji , Yang Yiwei , Zhang Jie , Men Kuo , Yi Junlin , Chen Ming , Dai Jianrong 

TITLE=Investigating transfer learning to improve the deep-learning-based segmentation of organs at risk among different medical centers for nasopharyngeal carcinoma

JOURNAL=Frontiers in Physics

VOLUME=11

YEAR=2023

URL=https://www.frontiersin.org/journals/physics/articles/10.3389/fphy.2023.1147900

DOI=10.3389/fphy.2023.1147900

ISSN=2296-424X

ABSTRACT=<p><bold>Purpose:</bold> Convolutional neural networks (CNNs) offer a promising approach to automating organ segmentation in radiotherapy. However, variations of segmentation protocols made by different medical centers may induce a well-trained CNN model in one center and may not perform well in other centers. In this study, we proposed a transfer learning method to improve the performance of deep-learning based segmentation models among different medical centers using nasopharyngeal cancer (NPC) data.</p><p><bold>Methods:</bold> The NPC data included 300 cases (S_Train) from one institution (the source center) and 60 cases from another (the target center), divided into a training set of 50 cases (T_Train) and a test set of 10 target cases (T_Test). A ResNet CNN architecture was developed with 103 layers. We first trained Model_S and Model_T from scratch with the datasets S_Train and T_train, respectively. Transfer learning was then used to train Model_ST by fine-tuning the last 10 layers of Model_S with images from T_Train. We also investigated the effect of the numbers of re-trained layers on the performance. The performance of each model was evaluated using the dice similarity coefficient, and it was used as the evaluation metrics. We compared the dice similarity coefficient value using the three different models (Model_S, Model_T, and Model_ST).</p><p><bold>Results:</bold> When Model_S, Model_T, and Model_ST were applied to the T_Test dataset, the transfer learning (Model_ST) had the best performance. Compared with Model_S, the <italic>p</italic>-values of all organs at risk were less than 0.05. Compared with Model_T, the <italic>p</italic>-values of most organs at risk were less than 0.05, but there was no significant statistical difference in Model_ST for the brain stem (<italic>p</italic> = 0.071), mandible (<italic>p</italic> = 0.177), left temporal lobes (<italic>p</italic> = 0.084), and right temporal lobes (<italic>p</italic> = 0.068). Although there was no statistical difference for these organs, the mean accuracy of Model_ST was higher than that of Model_T. The proposed transfer learning can reduce the training time by up to 33%.</p><p><bold>Conclusion:</bold> Transfer learning can improve organ segmentation for NPC by adapting a previously trained CNN model to a new image domain, reducing the training time and saving physicians from labeling a large number of contours.</p>