3D Deep Convolutional Neural Network for Mapping MRI to Synthetic MVCT for Radiotherapy Planning
J Scholey1*, A Rajagopal2, E Vasquez3, A Sudhyadhom4, P Larson5, (1) University of California San Francisco, San Francisco, CA, (2) University of California San Francisco, San Francisco, CA, (3) University Of California Berkeley, Berkeley, CA,(4) Harvard Medical School, Boston, MA, (5) University of California San Francisco, San Francisco, CA
Presentations
MO-IePD-TRACK 3-4 (Monday, 7/26/2021) 12:30 PM - 1:00 PM [Eastern Time (GMT-4)]
Purpose: To synthesize MVCT (sMVCT) datasets from MRI using a 3D deep CNN. This is a novel solution for MRI-only treatment planning where MRI can be utilized for anatomical delineation and MRI-generated sMVCTs for more accurate electron density mapping versus synthetic conventional kVCTs.
Methods: MVCT (3.5MV), kVCT (120kV), and T1-weighted MRIs for 120 head and neck cancer patients were retrospectively acquired and co-registered. MRI intensity values were normalized to standardize datasets acquired on different scanners. A deep neural network based on a fully-convolutional 3D residual U-Net architecture was implemented to map MRI intensity to sMVCT HU. Model inputs were volumetric patches generated from 100 paired MRI and MVCT datasets. The U-Net was initialized with random parameters and trained on a mean absolute error (MAE) objective function. Model accuracy was evaluated on 20 withheld test exams. For comparison, this method was applied to generate conventional synthetic kVCTs (skVCTs). sCTs were compared to respective actual CTs using MAE. To demonstrate clinical proof-of-concept, IMRT plans were generated on a MVCT then calculated with fixed plan parameters onto the corresponding sMVCT. Dose distributions were compared using gamma (3%/3mm local dose threshold) criteria.
Results: The model produced MAE of 118.4±17.3 and 127.7±33.3 HU for MVCT and kVCT datasets, respectively. Overall, there was good agreement between sCTs and CTs. Distinguishing bone from air posed the greatest challenges. The retrospective dataset introduced additional deviations due to sinus filling or tumor growth/shrinkage between scans, differences in external contours due to variability in patient positioning, or when immobilization devices were absent from diagnostic MRIs. Dose-volume-histograms were in close agreement between sMVCT and MVCT datasets and the gamma value between dose distributions was 100%.
Conclusion: sMVCT datasets can be generated from T1-weighted MRI using a 3D DCNN with dose calculation on a sample sMVCT in excellent agreement with the MVCT.
ePosters
Keywords
MRI, Megavoltage Imaging, Treatment Planning
Taxonomy
Not Applicable / None Entered.
Contact Email