Abstract
<title>Abstract</title> <p>Background: Automated cardiac magnetic resonance segmentation can fail after deployment outside development data, yet conventional validation usually reports average accuracy rather than whether an individual patient result should be accepted automatically or reviewed. We developed SHIFT-QA, a target-calibrated accept-or-review quality-assurance framework for medical image segmentation under pathology shift. Methods: SHIFT-QA combines ensemble-derived reliability evidence, including predictive uncertainty, mask disagreement, volume variability, and downstream-measurement instability, into a patient-level risk score. A labeled target-calibration subset is used to select an acceptance threshold before target-test labels are accessed. We implemented a local proof-of-concept on the Automated Cardiac Diagnosis Challenge (ACDC) cardiac magnetic resonance dataset using hypertrophic cardiomyopathy (HCM) as a pathologyheld-out target cohort. The audited cohort contained 100 patients and 200 end-diastolic/endsystolic records. Source training, source validation, target calibration, and target testing were patient-disjoint. Results: A lightweight two-dimensional U-Net reached a validation mean foreground Dice of 0.822 after continuation training. Failure-rule diagnosis showed that a strict minimum-class Dice rule was overly conservative for HCM calibration. The frozen operational rule defined failure as patient mean foreground Dice below 0.75 or absolute ejection-fraction error above 0.15. Target calibration selected a threshold of τ ∗ = 0.590787 with 80% calibration coverage. On the untouched HCM target-test set, SHIFT-QA accepted 8 of 10 patients and referred 2 of 10. The observed accepted-case failure rate was 0.000, whereas both referred cases were operational failures. Accepted patients had higher mean foreground Dice than referred patients (0.839 versus 0.657) and lower absolute ejection-fraction error (0.065 versus 0.228). Conclusions: This local ACDC prototype supports the feasibility of target-calibrated accept-or-review quality assurance for cardiac magnetic resonance segmentation under pathology shift. The evidence remains preliminary because the target-test set was small, the uncertainty estimate used a snapshot ensemble, and the final risk ranking used a predefined reliability-score fallback. Larger multi-centre and multi-vendor validation is required before clinical deployment. Trial registration: Not applicable.</p>