KEY POINTS
- This model-development study trained TUM-SAM on 301 patients with 2,548 brain metastases and externally evaluated it in 105 patients with 397 lesions. Three radiation oncologists participated in a blinded contour-preference study using a 20-patient subset.
- TUM-SAM combined nnU-Net lesion detection with tumor-adapted Med-SAM refinement to provide fully automated, prompt-free segmentation. Inference required less than 1 minute per patient.
- Against a single-physician reference, lesion-wise detection sensitivity was 94.0% with 0.89 false positives per scan. Against the union of two physician annotations, sensitivity was 85.6% with 0.35 false positives per scan, demonstrating sensitivity to reference definition.
- Relative to Physician A, TUM-SAM achieved a mean Dice similarity coefficient of 0.84 and a 95th-percentile Hausdorff distance of 1.9 mm, outperforming nnU-Net (0.69; 3.34 mm) and DeepMedic (0.68; 3.54 mm).
- Experts selected TUM-SAM contours in 81.2%–86.8% of raw comparisons. After adjustment for rater tendencies and case difficulty, estimated win probabilities remained 54.5%–56.0% versus each physician; errors persisted for very small and centrally necrotic, cystic, or hemorrhagic lesions.
CLINICAL TAKEAWAY
For clinicians planning stereotactic radiosurgery, this framework could reduce brain metastasis contouring workload while maintaining physician oversight. The findings are promising but not practice-changing because the preference study included only 20 patients, and prospective workflow, editing-time, treatment-planning, and clinical-outcome effects were not evaluated.