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

Area of Science:

• Medical imaging
• Machine learning
• Image-guided therapy

Background:

• Real-time tracking of anatomical targets is crucial for clinical interventions.
• Magnetic resonance-guided therapies often face challenges with 2D/pseudo-3D tracking and system latencies.

Purpose of the Study:

• To investigate methods for estimating real-time target displacement using 2D magnetic resonance images.
• To mitigate system latencies in magnetic resonance-guided therapies.

Main Methods:

• Utilized image registration to measure target displacements.
• Employed machine learning models (LightGBM, linear regression, LSTM, SVR) for incremental, patient-specific motion pattern prediction.
• Optimized hyperparameters and assessed models on independent datasets.

Main Results:

• Baseline registration error was 1.65 mm; system latencies introduced 0.81–1.98 mm additional error.
• Linear regression and SVR reduced latency-induced errors by 35–57%.
• Linear regression demonstrated high computational efficiency (15 ms).

Conclusions:

• Combining image registration with lightweight, patient-specific predictive modeling effectively mitigates system latencies for accurate real-time target tracking.
• Linear regression and SVR offer optimal accuracy and computational efficiency.
• Enables robust real-time tracking without motion assumptions.