Label-Efficient Semi-Supervised Cine-MRI Tumor Tracking Using a Segmentation Foundation Model
Abstract
Purpose
Accurate real-time tumor tracking is critical for MRI-guided radiotherapy, where geometric uncertainty can significantly increase dose to surrounding critical organs. Continuous cine-MRI enables motion-adaptive treatment. However, accurate tracking under large, non-rigid motion and heterogeneous image quality remains challenging. This study presents a personal prior-driven tumor tracking framework based on a fine-tuned foundation video segmentation model for real-time cine-MRI-based adaptive treatment.
Methods
We adapted and fine-tuned a video segmentation foundation model (SAM2) with a temporal memory mechanism for tumor tracking on the TrackRAD2025 cine-MRI dataset, comprising 585 2D sequences from six centers acquired on 0.35T and 1.5T MR-linac. The model propagates tumor masks sequentially across frames using the first-frame segmentation as a prompt, with memory attention enforcing temporal consistency. Architectural adaptations were introduced for single-channel MRI input and binary segmentation output. Training was performed on the labeled public subset (n=50) using a combination of segmentation losses and additional domain-specific losses targeting boundary accuracy and temporal smoothness. Performance was evaluated using Dice similarity coefficient (DSC), Hausdorff distance (HD), average surface distance (ASD), and center-of-mass distance (CD).
Results
On the evaluation set (n=50), the fine-tuned model achieved a mean DSC of 0.933±0.027, HD of 2.78±2.66 mm, ASD of 0.98±0.82 mm, and CD of 1.16±0.56 mm. Fine-tuning yielded consistent improvements over the pretrained model, increasing Dice from 0.914 to 0.933 and reducing HD from 3.43 to 2.78 mm. Compared with a baseline (U-net-based) tracking approach (DSC 0.772), the proposed method substantially reduced geometric and localization errors. Qualitative results demonstrated temporally stable mask propagation across challenging motion patterns.
Conclusion
We demonstrate a patient-specific tumor tracking approach that maintains accurate and temporally consistent localization under large, non-rigid motion in cine-MRI. By addressing a key practical limitation of current MR-linac workflows, this method supports reliable real-time motion management and strengthens the foundation for online adaptive radiotherapy.