Patient-specific automated multi-class anatomical motion tracking on real-time cine MR images using deep learning techniques.

backgroundReal-time multi-class motion tracking on cine images during magnetic resonance-guided radiation therapy (MRgRT) would enable instant monitoring of anatomical structures, facilitating dynamic beam control to minimize overdose to normal tissues and maximize tumor dose. However, conventional segmentation strategies incur significant time delays, making their clinical application impractical. Emerging deep learning (DL) technologies can provide a path to address this clinical challenge. PURPOSE: To propose a patient-specific framework for real-time automated simultaneous tracking of multiple anatomical structures on 2D magnetic resonance cine (cine MR) images acquired in two orthogonal planes, using DL techniques without the need for extensive training data.

methodsIn this study, we evaluated SegResNet, a DL-based auto-segmentation model, and TransMorph, a DL-based registration model. Additionally, conventional rigid and B-spline registration algorithms served as baseline models. Our datasets consist of 10 subjects, with both coronal and sagittal image sequences acquired for each subject using Elekta Unity MR-Linac protocols. With ground truth contours manually delineated for every frame, each model was trained and tested on a subject and plane-specific basis. Dice similarity coefficient (DSC), 95th-percentile Hausdorff distance (HD95), average surface distance (ASD), and center-of-mass errors (COME) were chosen to evaluate the output segmentations of each model. Furthermore, to analyze segmentation motion consistency across two orthogonal planes, the temporal rhythm and physical magnitude of the liver centroid trajectory were evaluated using the Pearson correlation coefficient (PCC) and the mean-centered Concordance correlation coefficient (CCC), respectively. Finally, a phase-resolved analysis was conducted to assess the robustness of each model across the motion cycle.

resultsSegResNet yielded significantly better results than registration models (p < 0.05). Averaged across all organs, planes, and subjects, SegResNet achieved an overall DSC of 0.97 ± 0.01, HD95 of 3.0 ± 0.9 mm, ASD of 1.1 ± 0.3 mm, and COME of 0.9 ± 0.4 mm. The PCC and mean-centered CCC of SegResNet are 0.77 ± 0.07 and 0.75 ± 0.08, respectively. Furthermore, phase-resolved analysis demonstrated SegResNet's phase-agnostic spatial stability, with no statistically significant differences (p > 0.05) observed across the three kinematic states for any organ across all evaluated segmentation metrics. The average inference time for SegResNet is approximately 0.3 ms per frame.

conclusionSegResNet outperformed TransMorph, as well as conventional B-spline and rigid registration models in terms of segmentation accuracy, cross-plane consistency, and robustness across different motion phases, and computation efficiency, potentially allowing for real-time simultaneous motion tracking of multiple anatomical structures on cine MR during MRgRT.