Main Session
Sep
28
PQA 03 - Digital Health Innovation and Informatics, Patient Safety & Quality, and Radiation and Cancer Biology
2586 - Evaluation of Computer Vision-Based Tracking Models for Tumor Localization in MR-guided Radiotherapy
Presenter(s)
Swati Rampalli, BS - University of Minnesota Medical School, Minneapolis, MN
S. Rampalli1, W. Choi2, K. E. Mooney2, N. L. Simone2, and H. Nourzadeh2; 1University of Minnesota Twin-Cities, Medical School, Minneapolis, MN, 2Dept. of Radiation Oncology, Sidney Kimmel Medical College and Comprehensive Cancer Center, Thomas Jefferson University, Philadelphia, PA
Purpose/Objective(s):
Magnetic Resonance Linear Accelerator (MR-Linac) systems enable real-time tumor visualization during radiation therapy, but reliable automated tracking remains essential to support adaptive workflows and potential dose escalation. General-purpose object tracking algorithms have shown promising speed and adaptability in visual environments (e.g., driving, surveillance), but have not been extensively validated in medical imaging. Prior efforts developed AI models trained specifically on medical datasets, but these often require extensive expert-labeled annotations and lack generalizability. We evaluated whether general-purpose object tracking models, originally developed for natural video data, can achieve clinically meaningful tracking accuracy on cine MRI without medical retraining.Materials/Methods:
SiamMask and XMem were evaluated on a subset of 50 patients from the TrackRad2025 Challenge Dataset across two MR-Linac field strengths: 0.35T (Institution A, n=25) and 1.5T (Institutions B and C, n=25). Each patient had 2D+t cine MRI sequences with expert-defined target contours. Models were initialized using the expert contour on the first frame and subsequently tracked the target across all frames. Predicted bounding boxes were compared with expert contours using Dice Similarity Coefficient (DSC), Intersection over Union (IoU), and 95th percentile Hausdorff Distance (HD95), reflecting overlap and boundary accuracy relevant to high-precision radiotherapy.Results:
Across all institutions, XMem demonstrated higher tracking accuracy than SiamMask. Mean DSC was 0.861 ± 0.168 versus 0.773 ± 0.232, mean IoU was 0.783 ± 0.183 versus 0.710 ± 0.244, and mean HD95 was 5.51 ± 7.02 mm versus 9.62 ± 8.30 mm, respectively. Both models showed reduced performance in 0.35T imaging, suggesting sensitivity to image quality and field strength. While overlap metrics approached clinically relevant thresholds in many cases, variability in boundary-based error was observed in a subset of sequences, driven by a small number of outlier frames with poor segmentation quality.Conclusion:
Off-the-shelf object tracking models demonstrate promising adaptability to cine MRI for MR-guided radiotherapy, with XMem outperforming SiamMask across descriptive metrics. While DICE and IoU values alone do not guarantee clinical viability as performance must also account for target complexity, proximity to critical structures, and steep dose gradients, these results offer a compelling baseline for applying general-purpose tracking architectures within MR-guided radiotherapy workflows. Further validation is warranted to assess clinical robustness in high-risk anatomical regions. Overall, this work supports continued development of automated tracking approaches to enhance treatment efficiency, targeted precision, and the potential for safe dose escalation in personalized radiation therapy.