Using Deep Learning Framework (pytorch) for Circular Cone Treatment Planning of CyberKnife System
B Liang1*, R Wei1, Y Li2, B Liu3,4, S Xu5, F Zhou3,4, Q Wu6, J Dai1, (1) National Cancer Center/National Clinical Research Center for Cancer/Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing, Beijing, CN, (2) Sun Yat-sen University Cancer Center, State Key Laboratory of Oncology in South China, Collaborative Innovation Center for Cancer Medicine, Guangdong Key Laboratory of Nasopharyngeal Carcinoma Diagnosis and Therapy, Guangzhou, Guangdong, CN, (3) Image Processing Center, Beihang University, Beijing, CN, (4) Beijing Advanced Innovation Center for Biomedical Engineering, Beihang University, Beijing, CN, (5) Chinese PLA General Hospital, Beijing, Beijing, CN, (6) Duke University Medical Center, Durham, NC
Presentations
TH-E-TRACK 6-7 (Thursday, 7/29/2021) 3:30 PM - 4:30 PM [Eastern Time (GMT-4)]
Purpose: With robot-controlled linac positioning, the CyberKnife system significantly increases the freedom in radiation beam placement, but also imposes more challenges on treatment plan optimization. This study proposed to use the deep learning framework (pytorch) for circular cone treatment planning of CyberKnife system.
Methods: The candidate beams were initialized by simulating the process of covering the entire target volume with equivalent beam taper. For each voxel, the total dose was the weighted summation of the dose contributed by all beams. This dose summation was topologized into a simple feed-forward neural network, and the optimization was converted into network training. The pytorch framework, capable of automatic differentiation and dynamic adjustment, was utilized to train the network. The least absolute shrinkage and selection operator (lasso) and group lasso regularization terms were utilized to reduce the number of beams and nodes, respectively.
Results: This method was validated and further compared with commercially available MultiPlan system on one typical lung case. This method protected the organs at risk (OARs) more effectively: the generated plan was of lower maximum dose within OARs. The homogeneity index (ratio of maximum dose to prescription dose, Dp) was reduced from 1.15 to 1.10. The volume covered by 50% (100%) Dp (V50 and V100) was reduced from 78.3cc to 59.3cc (14.6cc to 13.2cc), indicating more rapid dose fall-off. The number of nodes (beams) was reduced from 69 to 46 (146 to 93). In addition, the optimization time was shortened from around 30min to 52.3s.
Conclusion: In conclusion, the generated treatment plan achieved more satisfactory dose distribution with fewer nodes and beams. Besides, the optimization was 1-2 orders of magnitude faster than the MultiPlan system. Using the deep learning framework for treatment planning optimization has the potential to be used clinically.
Funding Support, Disclosures, and Conflict of Interest: The work was supported by the National Natural Science Foundation of China (11875320 and 81801799).
Handouts
Keywords
Optimization, Treatment Planning
Taxonomy
TH- External Beam- Photons: cyberknife
Contact Email