Size- and Burden-Stratified External Validation of Brain Metastasis Segmentation on 324 UCSF Imaging Studies

Abstract Background Independent institution-level validation of brain metastasis segmentation algorithms remains limited. We externally validated SPIRS-P1 and contextualized performance against prior algorithms and published inter-rater variability. Methods SPIRS-P1 was evaluated on 324 pre-treatment UCSF studies from 223 patients (3,449 reference-standard lesions). Per-case DSC, lesion-detection sensitivity, false positives, HD95, and NSD were evaluated by lesion size and burden. Native-space sensitivity and robustness analyses assessed dependence on preprocessing and lesion-matching definitions. Results Mean per-case DSC was 0.783 ± 0.155 (median 0.826), with 1.1 false positives per study. Sensitivity for micro, small, medium, and large lesions was 50%, 85%, 96%, and 99%, respectively; UCSF confidence intervals did not overlap Patel's internal results for micro and small lesions. Median study-level HD95 was 2.0 mm and NSD 0.936; among detected lesions, median HD95 was 1.0 mm and NSD 0.980. Native-space DSC was clinically similar to MNI-space DSC (median 0.848 vs 0.826). Conclusions SPIRS-P1 generalized well to independent UCSF data, with high sensitivity for lesions ≥125 mm³ and strong boundary agreement among detected lesions. Microlesion detection remained limited. Because the cohort contained only confirmed-positive scans, these findings support prospective evaluation as a contouring decision-support tool for known metastases but do not establish screening performance or deployment readiness.