Author: Claus Belka, Stefanie Corradini, George Dedes, Nikolaos Delopoulos, Christopher Kurz, Guillaume Landry, Ahmad Neishabouri, Domagoj Radonic, Adrian Thummerer, Niklas Wahl, Fan Xiao 👨🔬
Affiliation: Department of Radiation Oncology, LMU University Hospital, LMU Munich, Department of Medical Physics, LMU Munich, Department of Medical Physics in Radiation Oncology, German Cancer Research Center (DKFZ), National Center for Radiation Oncology (NCRO), Heidelberg Institute for Radiation Oncology (HIRO) 🌍
Purpose: In MR-guided online adaptive radiotherapy, MRI lacks tissue attenuation information necessary for accurate dose calculations. Instead of using deep learning methods to generate synthetic CT images from MRI, which could hinder real-time dose adaptation, we propose a fast novel deep learning-based dose calculation method directly on MRI and show its feasibility on prostate patient cases.
Methods: The 0.35 T planning MRI and deformed planning CT (registered to the planning MRI) of 30 prostate cancer patients treated with a 0.35 T MR-Linac were collected. All images were resampled to a voxel size of 1.5×1.5×1.5 mm³, and the air cavities (ACs) in the abdominal area of the deformed CT images were filled and corrected based on manual AC contouring on the MRI. Monte Carlo (MC) dose simulations under a 0.35 T magnetic field were performed on the corrected CT images. All photon beams were simulated using a uniform field size of 1×1 cm². 10800 beams were simulated with 5×106 initial photons for training (20 patients), 2160 beams with 5×107 photons for validation (4 patients), and 648 beams shooting through PTV with 5×107 photons for testing (6 patients). 3D MRI cuboids covering the photon beam were input into a Unet model to predict AC segmentation, then 3D MRI and predicted AC cuboids were input into an LSTM model for beam’s eye view processing to predict dose. The gamma passing rate γPR (2%/2mm, D> 10%Dmax), Dice similarity coefficient (DSC) of AC segmentation and beam dose profiles were evaluated.
Results: The mean γPR of test beams was 99.7% (range 92.1% to 100%). MC and predicted dose profiles matched well. The mean DSC of AC segmentation was 0.61. Total model prediction time was 12 ms/beam.
Conclusion: Our results indicate that the deep learning-based dose calculation directly on MRI is feasible for prostate cases.