Real-time target tracking using image registration and motion prediction.

purposeReal-time tracking of anatomical targets is critical in numerous clinical interventions. In magnetic resonance-guided therapies, targets are typically tracked only in 2D or pseudo-3D, and system latencies from image acquisition and processing introduce delays. This study investigates methods that leverage 2D magnetic resonance images to estimate the real-time displacement of a target from its reference position.

methodsTarget displacements from their reference position on a pre-treatment image were measured using image registration. These measurements served as inputs to machine learning models that capture patient-specific motion patterns and predict subsequent displacements. Specifically, the models were trained incrementally as measurements became available, adapting to patient-specific characteristics. LightGBM, linear regression, long short-term memory (LSTM) and support vector regression (SVR) models were evaluated for their ability to perform online updates from sparse measurements and for their accuracy in forecasting upcoming target displacements. Hyperparameters were optimized using five subject datasets representative of those acquired during clinical procedures. The methods were assessed on forty-nine independent datasets.

resultsTarget tracking achieved a baseline registration error of 1.65 mm. Introducing system latencies ranging from 200 to 600 ms resulted in additional latency-induced errors of 0.81 to 1.98 mm. Linear regression and SVR models provided the greatest accuracy gains, reducing the latency-induced errors by 35 to 57%. In contrast, the LightGBM and LSTM models exhibited comparatively lower predictive performance, likely due to limited training data and the increased model complexity. Linear regression was the most computationally efficient, requiring only 15 ms for combined online retraining and inference in both datasets.

conclusionsCombining image registration with lightweight, patient-specific predictive modeling to mitigate system latencies can enable accurate real-time target tracking. Among the evaluated methods, linear regression and SVR achieved the most favorable balance between prediction accuracy and computational efficiency. These findings support the feasibility of robust real-time tracking without reliance on prior assumptions about target shape or motion.