Pre-Deployment Evaluation of a Remote Service for Short-Axis Cardiac MRI Segmentation

Deploying cardiac segmentation models as remote inference services creates a black-box setting in which robustness under out-of-distribution shift is critical. We evaluated two candidate configurations for biventricular segmentation of short-axis cine cardiac magnetic resonance images trained on the same public benchmark dataset: a nnU-Net ensemble and nnSAM. Both were assessed within the same deployment-oriented framework on two external cohorts, a multi-site, multi-vendor adult dataset and a single-site, multi-scanner pediatric congenital heart disease dataset. Performance was evaluated using geometric metrics and biomarker-based endpoints, including clinically relevant ejection fraction (EF) failure rates. While performance was similar in the adult cohort, the nnU-Net ensemble was more robust in the pediatric cohort, with better geometric preservation and lower EF failure rates. These findings indicate that deployment-oriented validation of remote cardiac segmentation services should assess biomarker reliability alongside geometric accuracy.