(Sunday, 7/14/2019) 2:00 PM - 3:00 PM
Room: Stars at Night Ballroom 2-3
Purpose: The purpose of this study is to develop a robust CT image segmentation feature based on the existing U-Net deep-learning algorithms combined with the Inception module. The model is integrated in DeepPlan (a new commercial TPS) to improve the efficiency and accuracy of treatment plan design in radiotherapy.
Methods: An improved convolutional neural network was developed to classify the organs on CT images, which improves not only multi-organ segmentation accuracy but also runtime parallelization and task scheduling performance. For the classification network, the algorithm combines the results of 3 different feature extraction schemes to achieve accurate classification of organs after 4 convolutional operations. The purpose of organ classification is to determine the organs of each CT slice based on the classification results. Then, the CT image organ segmentation is carried out. Due to differences in the organ size and shape, different networks have been designed to yield more accurate organ boundaries. All networks are based on the U-Net and Iception modules. The convolution part of the original U-Net network is replaced by the Iception module which extracts more features, and the deconvolution part is replaced by the linear interpolation method. Several hundreds of CT image datasets are used as training data in this study. Manually segmentation by experienced radiologists are used as ground-truth. Keras framework is used to build network structure and training model.
Results: Thirty one organs are segmented automatically. Thirteen organs are found to have a higher Dice value than 90%, including the skin, lung and brain, and the Dice values of 9 organs were higher than 80% such as the thyroid, larynx, and esophagus.
Conclusion: A multi-organ automatic segmentation model has been developed and integrated in DeepPlan to provide a streamlined workflow for routine treatment plan design.
Funding Support, Disclosures, and Conflict of Interest: Supported by the National Key Research and Development Program of China (No.2017YFC0107500) and the National Natural Science Foundation of China (Grant No.11575180)
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
