Purpose: To account for respiration-induced motion during lung radiotherapy, indirect tracking requires surrogate-based correlations and system latency compensation. The present study proposes a unified Bayesian framework addressing both issues. Its potential to predict both future target motion from a surrogate signal and the related uncertainty is investigated.
Methods: The proposed framework is based on a Kalman filter theory specifically adapted for surrogate measurements. Since the quality of the surrogate motion affects the accuracy of the method, two external surrogates (abdominal and thoracic markers) and an internal surrogate (liver structure) are examined. Their motions and simultaneous lung target motion are collected for 9 volunteers during 4 minutes each, including drastic breathing changes to assess the robustness of the method. A comparison with an artificial non-linear neural network (NN) is performed. The first 15 seconds of each volunteer's sequence are used to train both methods, while testing is conducted under two scenarios: a worst-case scenario and a simplified one.
Results: Latencies ranging from 200 to 800 ms are tested. Surrogate predictability and target-surrogate correlation are also independently assessed to provide interpretable insights into the accuracy of the method. While the prediction errors from the thoracic surrogate are reduced with the NN in some cases, the proposed framework outperforms the NN in most cases when using the liver and the abdomen surrogates. The Bayesian method is found more robust to increasing latencies and to breathing variations. The thoracic marker is found to be less reliable than the other surrogates to predict the target position, whereas the liver appears as a better surrogate.
Conclusion: The proposed framework predicts both expected target positions and their associated confidence intervals. This study opens the way for real-time uncertainty prediction to further assist motion management decisions. Future work will investigate an adaptive version of the proposed framework.
Funding Support, Disclosures, and Conflict of Interest: This research was supported by Elekta, Medteq and NSERC (Collaborative Grant no. CRDPJ/502332-2016) and in in part by the Fonds quebecois de la recherche sur la nature et les technologies (FRQNT scholarship no. 271036).