Purpose: Cardiac toxicity is a common adverse side effect for patients undergoing radiation therapy of thoracic tumors. We hypothesized that a deformable registration model using deep learning (DL) could be trained to predict internal motion fields with sufficient resolution for monitoring cardiac motion.
Methods: A DL model for fast deformable image registration was trained using cine MRI scans acquired during MR-Linac treatments of thoracic and abdominal tumors. The model uses a pair of cine MRI images as inputs and outputs a motion vector field (MVF) which aligns the images. The trained model was applied to predict frame by frame motion from cine MRIs of test patients in which both cardiac and respiratory motion were visible. The number of respirations and heart beats in each cine were counted manually as well as automatically using peak detection on the high- and low-frequency components of the MVF displacements. Frequency analysis of MVF time series was performed to estimate the heart rate of each patient across multiple fractions.
Results: The DL model was trained using over 100,000 cine MRI frames from 86 treatment fractions. Average registration error (normalized RMSE) during five-fold cross validation was 0.032. Computation time required per registration was 8 ms. DL-based MVFs were then used to estimate respiratory and heart motion for 12 treatment fractions from 4 test patients. There was excellent correlation between DL-based and manually tracked respiratory and heart rates (Pearson’s correlation coefficient, respiratory, r(10)=0.90, p<0.001; heart, r(10)=0.85, p<0.001) with a mean absolute error of 0.50 breaths per minute and 2.1 beats per minute.
Conclusion: This is the first demonstration of simultaneous respiratory and heart motion monitoring using cine MRI from an MR-Linac. The method can likely be deployed to existing systems, enabling new beam gating strategies which better account for fast heart motions during radiation therapy.
Not Applicable / None Entered.