Answers are grounded in this paper and cite their sources.
Automated coronary artery segmentation from CT angiography (CCTA) can fail in ways that matter clinically, such as a missed proximal vessel, a broken trunk or a leak into a vein, while still reaching a high Dice score. We study label-free review triage: ranking reconstructions without ground truth so that reviewing a fixed 20% of scans catches as many anatomically critical failures as possible. We define a reference-based taxonomy of critical failures, pre-register an analysis plan with explicit success criteria, select one model on development data (a regularised logistic model over mask-derived graph, ostium and shape-plausibility features) and evaluate it once on a sealed test set of 160 patients (ImageCAS-X), segmented by a model it never saw. The frozen model reached a normalised recall of 0.906 at the 20% budget, catching 29 of the 32 critical failures that were catchable at that budget (29 of 64 in total) with precision 0.91. It met the pre-registered comparison with ensemble-agreement quality control. Its gain over the best pre-specified single-feature rule was +12.5 pp (95% CI -0.3 to +23.9), narrowly missing the pre-registered superiority criterion. In pre-registered follow-up analyses, the gain replicated across five nnU-Net folds (+11.4 pp, 95% CI +3.0 to +20.7; same patients), shrank or vanished when the input CTs were degraded, and was matched by an exploratory fragmentation feature. Simple, anatomy-aware topology features capture most of the achievable triage value; a learned combination adds a small, directionally consistent but fragile increment. We report every criterion, three negative findings and the full pre-registration trail, and release the code. Code: https://github.com/vhmaadhav/coronarysafe
No takes yet. Share an insight, caveat, or question.
Maadhav V H (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: