Predicting Future MRI Images for Real-Time Dose-Guided Adaptive Radiation Therapy
Abstract
Purpose
Magnetic resonance imaging (MRI) demonstrates strong potential to enable real-time imaging for dose-guided adaptive radiation therapy (ART). However, real-time ART with MRI is limited by long latencies, requiring temporal prediction to achieve accurate treatment adaptation. This study investigated the dosimetric accuracy of our patient-specific cross-attention future orthogonal planes (CAFOP) deep-learning framework to predict future images and enable latency-compensated real-time dose-guided ART.
Methods
Data from three liver cancer patients treated on a 1.5T Elekta Unity MRI-linac (48Gy in 3 fractions) were used to retrospectively to simulate real-time dose-guided ART via MLC tracking. Cine-MRI acquired in three orthogonal planes with 200ms intervals were used to train patient-specific models to predict future images across all three planes. CAFOP was used to generate cine-MRI with a prediction interval of 400ms and the liver was segmented on each image using the MedSAM2 foundational model. The segmentation centroid displacements were calculated, and these displacements were input into a dose-guided MLC tracking framework that adapted MLC apertures to minimise dosimetric deviations from the plan. Delivered doses were accumulated using the ground-truth motions and the dosimetric accuracy of CAFOP prediction was compared to prediction using a SwinLSTM model (a leading deep-learning framework for motion prediction), and when no motion prediction was used.
Results
The mean gamma-failure rates and standard deviations using a 2%/2mm criteria were lowest when tracking motion predicted using CAFOP (0.1%±0.1%), compared to tracking motion predicted using SwinLSTM (1.6%±2.6%), and when tracking without prediction (1.0%±2.2%). The mean gamma-failure rate was 9.9%±13.4% when motion was not tracked. The mean motion calculation time was 81ms for CAFOP, and 522ms for SwinLSTM.
Conclusion
While CAFOP and SwinLSTM can predict MRI images with improved image similarity compared to no prediction, CAFOP demonstrated the highest dosimetric accuracy, paving a pathway to enable latency-compensated real-time dose-guided ART.