Deep Learning-Based Automatic Plan Prediction for MR-Guided Adaptive Radiotherapy
Presentations
WE-IePD-TRACK 3-3 (Wednesday, 7/28/2021) 3:00 PM - 3:30 PM [Eastern Time (GMT-4)]
Purpose: Fast and automated plan generation is desirable for MR-guided adaptive radiation therapy (MRgART). Recently, significant progress has been made in predicting an optimal dose distribution using deep learning approaches. As a next step for automatic planning, we investigate the feasibility of predicting a deliverable plan for a target dose distribution based on deep learning.
Methods: A conditional generative adversarial network (cGAN) was trained to predict the MLC leaf sequence corresponding to a target dose distribution. The training data were 50 ground truth dose distributions and corresponding IMRT plans that were optimized during MRgART in a treatment planning system (TPS) for 10 pancreatic cancer patients (each with 5 fractions). The model input was the dose distribution from each individual beam and the output was the predicted corresponding field segments with specific shape and weight. The predicted plan was then generated and transferred back to TPS to calculate the corresponding dose. Patient-based leave-one-out-cross-validation was employed and the model performance was evaluated by calculating the gamma passing rate of the ground truth dose vs. the dose from the predicted plan.
Results: The average gamma passing rate (95%, 3mm/3%) among 10 test cases was 88%. With relaxed passing criteria (95%, 5mm/5%), the average passing rate was 96%. In general, we observed 95% of the prescription dose to PTV achieved with a slight increase of dose to OARs. Complete plans were predicted in <15 seconds using a GTX 1660TI GPU .
Conclusion: We have developed a deep learning method to predict IMRT plans rapidly and automatically for pancreatic cancer. With further developments using large datasets, the method may be implemented to accelerate MRgART process or to facilitate real-time MRgART.
Funding Support, Disclosures, and Conflict of Interest: This work is supported by Manteia Technologies.
ePosters
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email