Deep-Learning (DL)–Based Real-Time Inverse Planning for MR-Guided Radiotherapy (MRgART)
Abstract
Purpose
Treatment planning for MR-guided adaptive radiotherapy (MRgART) requires extensive time and effort in both preplanning and online adaptation processes. It is a major bottleneck hindering the efficiency and quality of MRgART. Specifically, extended preplanning process negatively impacts clinical outcomes of radiotherapy while prolonged online planning can introduce anatomical changes, diminishing the benefits of adaptation. To address this bottleneck, we proposed TransFM, a deep learning (DL)–based domain transformation framework for ultra-fast inverse plan optimization. By integrating with the in-house developed dose prediction and dose calculation models, this framework enables an end-to-end real-time planning workflow for MRgART.
Methods
The main idea of TransFM is to approximate inverse plan optimization using a deep neural network (DNN). To efficiently model this complex high-dimensional transformation from dose and image volumes to fluence maps (FMs), TransFM employs an innovative shuffling-convolution scheme to effectively approximate global operators via cost-effective convolutions. A binary beam-angle embedding is also constructed to encode beam configuration into the model. TransFM was trained with a real-time dose estimation model in a unified framework to take advantage of randomly generated FM–dose pairs as additional training samples. A cohort of 209 prostate cases with 1,133 initial and online ART plans was collected. 177 cases (866 plans) and 17 cases (130 plans) were randomly selected for training and validation, while the rest were saved for independent testing.
Results
TransFM can estimate FMs based on the input image volume, beam configurations, and desired dose distribution within 30ms for each case. The dose calculated using the FMs estimated by TransFM achieved γ-passing rate (3%/2mm) of 98.1%±2.82% compared to the desired clinical dose, with mean dose difference <1.5Gy for both Dmax and Dmean across all structures.
Conclusion
TransFM enables ultra-fast inverse planning for both preplan and online adaptation, demonstrating strong potential to support real-time treatment planning for MRgART.