Markerless Liver Tumor Tracking during Radiotherapy Using Segment-Anything Model Based on Radiopaque Contrast Agent
Abstract
Purpose
Targeting accuracy in stereotactic ablative body radiotherapy (SABR) for hepatocellular carcinoma (HCC) is often hindered by the poor soft-tissue visibility of tumors on standard image-guided radiotherapy using kilovoltage (kV) images. However, ~60% of HCC patients have previously undergone transarterial chemoembolization (TACE), leaving residual radio-opaque lipiodol near the tumor site. To overcome current imaging limitations, we exploit lipiodol as an intrinsic contrast agent to achieve accurate markerless tracking on real-time kV images using the Medical Segment Anything Model (Med-SAM).
Methods
This workflow leverages 2D lipiodol contours from the planning CT as spatial prompts for a tailored Med-SAM architecture. Given the high positional reproducibility of breath-hold maneuvers, these planning-derived prompts provide a reliable surrogate for real-time target position with minor adjustment accounting for intrafraction motion. We fine-tuned Med-SAM on a patient-specific basis using digital reconstructed radiographs (DRRs) simulated from planning CTs. To bridge the domain gap between DRRs and real-time kV images, a dual-stage augmentation flow was implemented: 3D lipiodol contrast manipulation and 2D noise and intensity perturbations. Tracking errors were measured as centroid distance relative to two expert annotations.
Results
Five of 14 patients across multiple institutions have kV-visible lipiodol for labelling. A total of 135 X-ray projections from 11 fractions were analyzed. Med-SAM enabled real-time inference with mean tracking errors of −0.2 ± 2.1 mm (SI) and 0.3 ± 1.2 mm (LR/AP), with 89.6%/96.3% (SI) and 96.3%/100% (LR/AP) of frames within 3/5 mm, respectively. For reference, interobserver variability was 0.0 ± 2.1 mm (SI) (88.9%/95.6% within 3/5 mm) and 0.2 ± 1.6 mm (LR/AP) (94.8%/97.8% within 3/5 mm).
Conclusion
By leveraging existing planning data and a tailored foundation model, this patient-specific workflow achieves expert-level accuracy that meets clinical error tolerances. This approach demonstrates strong potential for clinical deployment to improve treatment delivery without the need for implanted fiducials.