A Novel Leaf Sequencing Optimization Method with Deep Reinforcement Learning
W Zhang1*, L Kong1, Q Zhou1, S Xu2, (1) Manteia Medical Technologies, Xiamen,CN, (2) Chinese PLA General Hospital, Beijing, CN
Presentations
PO-GePV-E-22 (Sunday, 7/25/2021) [Eastern Time (GMT-4)]
Purpose: Due to the machine limitations inherent in the design of multi-leaf collimator (MLC), the current leaf sequence algorithm has problems such as insufficient execution efficiency and delivery accuracy. This study designed a leaf sequencing method based on reinforcement learning, trying to solve the above problems.
Methods: The training engine consists of MLC simulation environment, data generator and discrete action network structure. First, a simulated fluence map was obtained through the data generator, which, together with the current collimator position and the cumulative conformal intensity map, was entered into agent network. The predicted collimator movement and monitor units was applied to environment, which made a new collimator state and intensity map. The reward was calculated with the final target and was back-propagated to complete a whole cycle. The design of MLC simulation environment would consider machine parameters and constraints, which greatly reduced the action space of discrete action model. The Bezier curve and elastic transformation were innovatively used in the data generator to simulate the intensity distribution, making the training target closer to the actual fluence map. Using nasopharyngeal carcinoma IMRT as a test bench, we tested and compared the accuracy difference between the proposed and traditional leaf sequencing algorithms.
Results: During test using fluence map extracted from 20 IMRT plans of nasopharyngeal carcinoma, with an intensity level of 64, the proposed method achieved a relative intensity difference of 1.25% with 15 segments, while the traditional algorithm requires 40 segments to achieve 1.58%. When the strength level increased to 128, the method achieved 0.78% with 15 segments, while the traditional algorithm required 50 segments to achieve 1.06%.
Conclusion: Our novel leaf-sequencing method based on reinforcement learning has obvious advantages in execution efficiency and delivery accuracy. Experiments shows that this method has great potential in the field of MLC leaf-sequencing.
ePosters
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email