Risk Control Certifies Rates, Not Magnitudes: An Audit of Selective Fetal Biometry

Risk Control Certifies Rates, Not Magnitudes: An Audit of Selective Fetal Biometry

In their paper, Zhengshuai Yang, Linna Jiang, Changyun Li and Guihu Zhao argue that risk control can limit how often a selective prediction system makes gross errors without limiting how large an individual error may be. They audit a prenatal ultrasound pipeline that estimates head circumference (HC) and can refuse to answer, using two separately pre-registered splits of 999 labelled HC18 images. On both splits, the pipeline produced a large error despite a high contour-completeness score; the authors say one outlier among approximately 150 answers left the certified risk bound intact. Their central point is that the score and certificate address error frequency, not whether a particular accepted measurement is anatomically plausible. The authors report that raising the lower bound of the admissibility region from 50 to 55 mm cut worst-case error on both splits, at a cost of 0.4 coverage points; suppressing the error through the completeness threshold cost substantially more coverage. At an anatomically defined 70 mm, frozen tests gave 1.91 and 1.81 mm MAE. The results describe this audit, not evidence that the pipeline is reliable in clinical use.

Publication: 811cf3e6-5322-40a7-81ba-88685cc947c0

attribution — pass: The reading correctly attributes the paper and its argument to Zhengshuai Yang, Linna Jiang, Changyun Li and Guihu Zhao, matching the arXiv author list. Results are framed as the authors' reports, not established fact.

Authors:Zhengshuai Yang, Linna Jiang, Changyun Li, Guihu Zhao

summary — uncertain: The summary keeps the paper's thesis but omits the failure's nature: a small, complete ring was fitted where the true HC was 324.6 mm, so the output was implausibly small. Without this, readers cannot see why raising the lower bound from 50 to 55 mm fixes it. Central terms are also unexplained: risk control, coverage, admissibility region (the source calls it the anatomical range where the measurement is defined) and MAE. Smallest repair: one clause on the too-small ring, plus plain glosses or tooltips for coverage (share of images answered), admissibility region and MAE.

a small, geometrically complete ring fitted to an image whose true HC is 324.6 mm

date — pass: originalPublished 2026-11-04 matches the arXiv v1 submission date. The later feed date is correctly not used.

[Submitted on 4 Nov 2026]

support — pass: The claims match the abstract: the rate-not-magnitude bound, the 0.94 completeness score on the failing case, replication on the second split, the 0.4-point coverage cost, worst-case errors of 11.4 and 11.1 mm, and 1.91 and 1.81 mm MAE. The clinical-use caveat and generalization question are appropriately limited.

A second, untouched split reproduces the mechanism on a different image (233.3 mm).

media — pass: The selected media is the verified arXiv abstract page for this paper. It is the direct original source.