Real-time tumor tracking for magnetic resonance-guided radiotherapy using label-efficient foundation model adaptation

Why clinicians should know about it

• Picked for Medical Physics (top studies of the week, 20 September 2026): MedSAM2 adaptation enables real‑time cine‑MRI tumor tracking

Abstract

BACKGROUND AND PURPOSE: Real-time tumor tracking on cine magnetic resonance imaging (cine-MRI) enables intra-fraction motion management in magnetic resonance-guided radiotherapy. Pipelines often rely on deformable registration or template matching to propagate contours, which can degrade under non-rigid motion and cine-MRI artifacts. This study evaluated two MedSAM2-based adaptations for real-time tracking with limited labels. MATERIALS AND METHODS: The TrackRAD2025 dataset includes sagittal cine-MRI sequences from six institutions acquired on 0.35 T and 1.5 T MRI-linear accelerators, with 50 labeled and 477 unlabeled training cases and 50 test cases. Tracking was formulated as frame-wise target segmentation from the first-frame ground-truth mask. Method A fine-tuned MedSAM2 on 40 labeled cases and used rank-weighted checkpoint averaging. Method B combined two MedSAM2 models trained on different labeled splits augmented with pseudo-labels, and merged their predictions during inference. RESULTS: On the hidden test set, Method A ranked highest on Dice similarity coefficient (DSC), center distance (CD), 95th-percentile Hausdorff distance (HD95), and mean average surface distance (MASD), with DSC = 0.891, CD = 1.47 mm, HD95 = 4.22 mm, MASD = 1.66 mm, relative dose to 98% of the target volume ( D 98 % ) = 0.936, and runtime of 0.04 s per frame. Method B achieved similar geometric performance and numerically higher relative D 98 % (0.953), at 0.10 s per frame. Paired Wilcoxon signed-rank testing showed no significant differences after correction for multiple testing. CONCLUSIONS: MedSAM2 adaptation enabled real-time cine-MRI tumor tracking with limited labels. The single-model approach provided faster inference and simpler deployment, while semi-supervised ensembling showed no significant improvement at higher computational cost.

Abstract as published, via PubMed.