Feasibility of Predicting Deliverable Daily Plans Using Deep Learning for MR-Guided Adaptive Radiation Therapy
S Hamdan1*, H Nasief1, L Buchanan2, Y Zhang1, E Omari1, X Li1, (1) Medical College of Wisconsin, Milwaukee, WI, (2) Brown University, Providence, RI
Presentations
SU-H400-IePD-F8-2 (Sunday, 7/10/2022) 4:00 PM - 4:30 PM [Eastern Time (GMT-4)]
Exhibit Hall | Forum 8
Purpose: Fast and automated generation of daily adaptive plans is desirable in MR-guided adaptive radiotherapy (MRgART). Building up from our previous studies, this work aims to investigate the feasibility of using deep learning to quickly predict deliverable daily adaptive plans based on a desirable dose distribution of a daily MRI for MRgART.
Methods: A total of 145 daily IMRT plans (ground truth) created for 145 daily (fraction) MRIs acquired during MRgART of 24 pancreatic cancer patients on a 1.5T MR-Linac were decomposed into 1,635 individual beams. For each beam, the cumulative dose distribution of all IMRT segments was projected onto the beam's eye view (BEV) plane, and the beam parameters, i.e., MU numbers, leaf positions, and diaphragm positions were represented pictorially by concatenating the segment apertures weighted with their corresponding MU numbers in a NumPy Array. A conditional generative adversarial network (cGAN) was trained using 1,580 beams from 140 fractions, with 5 fractions of one patient, comprising of 55 beams, left out for validation purpose. The obtained model required the input of the desirable BEV dose distribution and output the pictorial representation of the beam parameters. The predicted dose distribution for each beam was generated using a conventional treatment planning system (TPS) based on the predicted beam parameters and was compared to the desirable dose distribution based on gamma analysis (5% threshold, 5mm/5%).
Results: Gamma analysis showed the maximum passing rate of 95.6% and mean of 93.7%. The time for a beam prediction was less than 1 second using a GTX1660TI GPU.
Conclusion: It is feasible to use deep learning to quickly and automatically generate daily deliverable adaptive plan based on the desirable dose distribution on the daily MRI. With further developments of using large datasets and inclusion of patient anatomy, the method may be implemented to accelerate MRgART.
Funding Support, Disclosures, and Conflict of Interest: This work is partially supported by Froedtert and Medical College of Wisconsin Cancer Center Foundation and the Medical College of Wisconsin Fotsch Foundation.
Keywords
Image-guided Therapy, Radiation Therapy, MR
Taxonomy
IM/TH- MRI in Radiation Therapy: Development (new technology and techniques)
Contact Email