Analysis of Real-Time Cardiorespiratory Motion Prediction Algorithms for Mrigrt Stereotactic Arrhythmia Radioablation
Abstract
Purpose
Ventricular tachycardia (VT) arises when abnormal electrical impulses create rapid, self-sustaining ventricular activation instead of normal sinus rhythms. Stereotactic arrhythmia radioablation (STAR) is increasingly used as salvage treatment in refractory VT patients. Magnetic resonance–guided radiotherapy with multi-leaf collimator (MLC) tracking under cine-MRI guidance for STAR is expected to further enhance precision and reduce side effects. However, predicting combined cardiorespiratory motion is challenging for classical methods such as linear regression (LR), making system latency an unresolved issue.
Methods
We have collected multi-institutional cine-MRI (8-11 Hz) of the left ventricle and extracted cardiorespiratory motion traces totaling 190min using Segment Anything Model 2 (SAM2). The segmentation accuracy of SAM2 was evaluated against manual segmentations from 4 observers. Long-short-term-memory networks (LSTMs) trained on synthetic and cine-MRI cardiorespiratory motion traces, assuming 360ms MRI-linac latency, were compared to LR on four independent test sets. Both methods were updated online using the most recent data. Next, we evaluated how each prediction algorithm performed on deconvolved cardiorespiratory motion components, obtained through frequency filtering. Root-mean-square-error (RMSE) and the Wilcoxon signed-rank test were used as metric and statistical test.
Results
The RMSE between the segmentation center-of-mass of SAM2 against observer 1 (2.7±2.0 mm) and the average observer variation (3.4±1.5 mm) were comparable (p=0.14). The online LSTM (median±IQR RMSE: 2.4±1.1mm) outperformed LR (2.8±1.2mm) for combined motion (p<0.0001). The LSTM clearly outperformed the LR for respiratory motion (1.3±0.5mm vs. 2.0±1.0mm, p<0.0001). Both models performed comparably for cardiac motion (1.8±0.8mm vs. 1.9±0.6mm, p=0.98), suggesting that LR failed to predict the low-frequency motion component. Inference times of 5±1ms and online retraining times of 34±12ms (LSTM), 4±1ms (LR), and 49ms (SAM2) were measured.
Conclusion
We demonstrated that state-of-the-art online LSTMs can model cardiorespiratory motion simultaneously, which LR did not achieve. Online LSTMs exhibited a prediction performance comparable to the segmentation accuracy of SAM2.