Efficient real-time 3D tracking of liver targets through image registration and LightGBM
Ayaz Nakhuda, HaPhan Tran, Elodie Lugez · 2025
Background: The ability to track regions of interest in real time is essential for many clinical procedures. However, many systems are constrained to 2D tracking and experience delays due to image acquisition and processing times. This research presents innovative techniques to estimate 3D target movements in real time by utilizing interleaved coronal and sagittal magnetic resonance images. Methods: Image registration is employed to quantify the target’s 2D movement from its original position on a reference image. This 2D information is then combined with predictions from LightGBM models to determine the entire 3D displacement of the target. Additionally, the 2D displacement measurements are incorporated into the LightGBM models’ training set for continuous on-line re-optimization. The methods were evaluated using a curated dataset of real liver data; their performance in tracking and ability to offset system delays was analyzed. Results: On average, the image registration method yielded tracking errors of 1.19 mm. System delays of 200 ms, 400 ms and 600 ms led to tracking errors of 1.78 mm, 2.62 mm and 3.46 mm. The trained LightGBM models, once trained, reduced these errors by 29% to 46%. Conclusions: The challenges of incomplete data and system delay were effectively addressed by the investigated methods, which demonstrate great potential for real-time 3D tracking of various targets without needing prior knowledge of their displacement.