Author: Steve B. Jiang, Dan Nguyen, Chenyang Shen, Fan-Chi F. Su, Jiacheng Xie, Shunyu Yan, Daniel Yang, Ying Zhang, You Zhang 👨🔬
Affiliation: Medical Artificial Intelligence and Automation (MAIA) Lab, Department of Radiation Oncology, UT Southwestern Medical Center, UT Southwestern Medical Center, Medical Artificial Intelligence and Automation (MAIA) Lab & Department of Radiation Oncology, UT Southwestern Medical Center, Medical Artificial Intelligence and Automation (MAIA) Laboratory, Department of Radiation Oncology, UT Southwestern Medical Center, The University of Texas at Dallas 🌍
Purpose: Fast dose verification is essential for the safety and efficiency of MR-guided adaptive radiotherapy (ART) as patients anxiously waiting on the treatment couch. Conventional tools often require several minutes minimally, hindering efficiency and quality of ART. This study aims to apply a novel geometry-encoded deep-learning (GeoDL) framework to, for the first time, achieve real-time 3D dose verification for MR-guided ART.
Methods: GeoDL framework integrates fluence maps (FMs) and electron density (ED) maps by explicitly encoding the exact treatment geometry on MR-guided linear accelerator (MR-LINAC). Specifically, FM of each beam was constructed based on the optimized multi-leaf collimator apertures and their corresponding monitor units. It was then integrated with ED volume, embedding FM into the image domain to mitigate the challenging domain transformation. Specifically, each FM was projected onto a plane 30cm away from isocenter, preserving beam delivery geometry while avoiding potential overlap with patient body and motion management device. A 3D U-Net was subsequently trained to directly estimate 3D dose distribution based on the FM-encoded ED volume. We collected a cohort of 209 prostate cases with 1,133 initial and online ART plans calculated using Monte-Carlo algorithm in TPS. 177 cases (866 plans) and 17 cases (130 plans) were randomly selected for training and validation, respectively, while the rest were saved for independent testing.
Results: The end-to-end dose verification process, including geometry encoding and model evaluation, averaged ~31.8ms per patient. For testing cases, the average γ-passing rate (3%/2mm) was 98.4%±3.54%, with 0.84%±1.14% error in Dmean and 1.50%±1.05% in Dmax for treatment targets. Dose estimation errors in critical organs were consistently <1.5%.
Conclusion: This study establishes the GeoDL framework as a first-of-its-kind real-time 3D dose verification tool for MR-guided ART, delivering sufficient accuracy as secondary dose verification with unmatched speed, offering a significant advancement in efficiency and quality of the workflow.