Deep Learning-Based Synthetic CT Generation for MR-Only Adaptive Radiation Therapy On MR-Linacs
B Hunt1, B Zaki2, G Russo3, G Asher4*, G Gill5, C Thomas6, T Prioleau7, D Gladstone8, B Pogue9, R Zhang10, (1) Dartmouth College, Hanover, NH, (2) Dartmouth-hitchcock Medical Center, ,,(3) Dartmouth-hitchcock Medical Center, ,,(4) Dartmouth College, ,,(5) Dartmouth-hitchcock Medical Center, ,,(6) Dartmouth-Hitchcock Medical Center, ,,(7) Dartmouth College, ,,(8) Dartmouth College, Hanover, New Hampshire, (9) University of Wisconsin-Madison, Madison, WI, (10) Dartmouth-Hitchcock Med. Ctr., Lebanon, NH
Presentations
PO-GePV-M-240 (Sunday, 7/10/2022) [Eastern Time (GMT-4)]
ePoster Forums
Purpose: In existing MR-guided on-table adaptive radiation therapy (MRgRT) workflows, electron density of the target and organs-at-risk are obtained by deforming a simulation CT scan to the current MR setup scan. CT acquisition adds significant time to MRgRT workflows and imaging dose to the patient. Registrations have inherent limitations which frequently require manual corrections and is prone to human error. We hypothesize that synthetically generated CTs (SynCTs) via deep learning could be utilized to accurately predict electron density for subsequent adaptive planning and dose computation.
Methods: Simulation MR/CT scans from 20 patients undergoing MRgRT were used to train a generative encoder-decoder network that predicts SynCTs from MR setup scans. All patients underwent a standard CT simulation followed by an MR simulation scan on the MR-Linac approximately 1 hour later. The trained network then predicted electron density for a different setup scan acquired at subsequent treatment fractions. These SynCT volumes were directly compared to the deformed CT volumes using mean absolute error in Hounsfield units (HU). SynCT volumes were also exported as DICOM sequences and re-imported to the MR-Linac treatment planning system to recalculate dose and assess differences in resulting dose distributions, dose-volume histograms, and clinical dose prescriptions.
Results: The trained network achieved good accuracy on both training and testing datasets (mean absolute error in HU ± standard deviation, train: 38.4 ± 19.1 HU, test: 45.9 ± 24.5 HU). The average time required to generate a SynCT for a single MR volume using the trained model was 0.63 ± 0.06 seconds. Using the SynCT volumes for dose re-optimization did not compromise the quality in all 20 plans.
Conclusion: Accuracy of deep learning-based synthetic CT generation using low-field MR setup scans on MR-Linacs was sufficient for accurate dose calculation/optimization. This approach can enable MR-only treatment planning workflows on MR-Linacs.
Keywords
MRI, CT, Stereotactic Radiosurgery
Taxonomy
IM/TH- MRI in Radiation Therapy: Synthetic CT
Contact Email