Purpose: To compensate for the latency in real-time MR-guided radiotherapy, accurate forecasting of tumor motion during the delivery is critical for MLC-based tumor tracking. The purpose of this study is to investigate the feasibility and accuracy of using an RNN model, trained online before delivery, to predict tumor intrafraction motion.
Methods: MRI-Linac comes with the capability of simultaneous cine-imaging during delivery. The system latency of the MR-Linac was about 200-500ms,including the time of MR-imaging, target detection, and MLC motion. A single long-short-term-memory (LSTM) layer with 500hidden units was used to construct the prediction RNN. Tumor motion in superior-inferior direction obtained from ten lung-cancer patients with various tumor location were utilized retrospectively for evaluation. Two training length of 30 and60sec were evaluated with four prediction horizons of 200, 300, 400, and500mSec. The impact of imaging frequency was also tested for 10 and30Hz. The performance of the motion forecasting is evaluated in term of root-mean-square-error(RMSE) and factors that impact the accuracy were assessed.
Results: The RNN-network can be trained in 40sec and 80sec with 30sec/10Hz and 60sec/30Hzc data, respectively. The mean RMSE of ten patients was 0.98Â±0.36mm, 1.22Â±0.56mm,1.61Â±0.83mm, and 2.09Â±1.21mm for prediction horizon of 200, 300, 400, and500ms respectively, with 30sec of 10Hz training data. Increasing training length to 60sec or imaging frequency to 30Hz doesn't make significant improvement to the forecasting accuracy. The accuracy varied significantly with patients and was strongly correlated to the variation of tumor excursion, with correlation coefficient of 0.83(P<0.01). Using 4mm as threshold of excursion STD can select a subgroup of 7patients with<1mm mean RMSE at 300ms and<1.5mm with 500ms prediction window.
Conclusion: An RNN network can be trained online with 30sec motion data and perform accurate tumor motion forecasting. The algorithm can be utilized to compensate for the system latency of MR-Linac for real-time tumor tracking.
Funding Support, Disclosures, and Conflict of Interest: This research was partially funded by an Elekta research grant.