Gantry Motion Artifact Reduction Using Deep Learning: Towards Volumetric Modulated Arc Therapy On MR-Linacs
B Hunt1*, Z Han1, BI Zaki2, GA Russo2, BW Pogue1, DJ Gladstone2, R Zhang1,2, (1) Thayer School of Engineering, Dartmouth College, Hanover, NH (2) Dartmouth Hitchcock Medical Center, Lebanon, NH
Presentations
SA-B-Therapy Room-3 (Saturday, 4/17/2021) 12:30 PM - 2:30 PM [Eastern Time (GMT-4)]
Purpose: Current MR-Linacs use MR-guided beam gating at fixed delivery angles. Gantry motion introduces significant image artifacts and thereby inhibits real-time target tracking. This work evaluates a novel deep learning-based approach for gantry motion artifact reduction using clinical MR-cine data acquired during routine clinical use of an MR-Linac.
Methods: MR-cine image sequences from patients undergoing radiotherapy were accessed under an IRB approved protocol. A quantitative metric to detect gantry motion artifacts was developed and utilized to partition the dataset into artifact and artifact-free image sequences, with two-thirds used for training and one-third for testing. A previously developed deep neural network model (known as cycleGAN) was trained to synthesize artifact-free images from those containing artifacts and vice versa. Model predictions for artifact removal were then assessed on the test dataset.
Results: A total of 1,664 cine images from five patients were included in the analysis (train: 653 artifact and artifact-free each, test: 358 artifact). Artifact detection using frame-by-frame MSE was robust metric to partition artifact-free and artifact data in all patients. Using the network model on artifact frames in the test set resulted in an increase in both structured similarity index measure and peak signal-to-noise (paired t-test, SSIM: 0.83 vs 0.72, p<0.01; PSNR: 24.5 vs 21.3, p<0.01). Performance on 135 frames from a patient withheld from training was comparable to average performance (SSIM: 0.83, PSNR: 25.3), indicating robustness against patient specific anatomy. Computation time for processing of a single image was less than 20 milliseconds in deployment.
Conclusion: This study is the first to propose a computationally efficient solution for reduction of gantry motion artifacts on MR-Linacs and could help enable real-time target tracking for volumetric modulated arc therapy. Future studies should evaluate the clinical utility of the approach and its ability to support real-time target tracking during gantry motion.
Keywords
Low-field MRI, Image Artifacts, Targeted Radiotherapy
Taxonomy
IM/TH- MRI in Radiation Therapy: MRI/Linear accelerator combined- IGRT and tracking
Contact Email