Automated labour ultrasound angles sometimes miss by tens of degrees: referring the most suspicious 5% cuts gross errors from 1.2% to 0.2%

Automated AoP measurement can now carry a reliability score for each frame that needs no reference label, so the most suspicious fraction can be sent for manual review instead of accepting or repeating everything. Such automated measurements were accurate to a few degrees on average, but the error distribution had a long tail in which a few measurements were wrong by tens of degrees, and nothing in the output marked them. On the 501 IUGC2025 test files, the deployed six-detector median prediction had a mean error of 4.615 degrees, with 6 files (1.2%) above 20 degrees and 47 (9.4%) above 10 degrees.

Ranking by disagreement and referring 5% of frames lowered the residual rate of errors above 20 degrees to 0.2%, against 1.2% under random referral. Referring the same fraction at random leaves the residual rate of errors above 20 degrees at its analytical expectation of 1.2%. At the 5% budget of 25 frames, angle-space and coordinate disagreement each caught five of six such errors, heatmap confidence and the ensemble mean heatmap peak four each, and single-model input consistency three; 20 of the 25 referred frames were not failures by this definition.

The reliability scores are computed from the predictions of six heatmap-based landmark detectors alone, without reference annotations. The candidate scores measure different things: disagreement among detectors, a single detector's consistency under input perturbations, heatmap peak values, and the peak of the averaged ensemble heatmap. The pool holds three HRNet-W48 and three ConvNeXt-V2 detectors, and the deployed prediction is the coordinate-wise median of the six landmark configurations, from which the angle is computed; on the held-out Test2 split of ISBI 2015 cephalometric data, a three-detector ensemble referring the 10% most-disagreed images lowered retained mean radial error by 0.281 mm (95% CI 0.135 to 0.460).

The results are frame-level analyses of the IUGC2025 test split: 501 files and 493 distinct images, with two file-name prefix groups and no patient mapping, so no patient-level statement is possible. The failure thresholds of 10, 15 and 20 degrees are research definitions not validated against labour outcomes. The cephalometric replication uses the ISBI 2015 Test2 split, a task without a comparable tail of gross failures. The intended readers are researchers and engineers working on medical image quality control and selective prediction.

The tail above 20 degrees holds only six cases, so counts at that threshold are fragile. Three failures above 20 degrees exceed the threshold by less than 2.5 degrees, while byte-identical duplicate images differ in annotation by up to 5.3 degrees, so which frames count as gross failures depends on the reference. The validation split overlaps the unlabelled release used for self-training, which may have made one detector's validation-based selection optimistic. Angle-space disagreement flagged 11.2% of test frames at a nominal 5%, and resampling the 100 calibration frames alone produced rates from 3.8% to 14.6%. Only three of the twelve frames with a reference above 140 degrees were referred at the 5% budget, and this post hoc analysis needs confirmation in independent data.