False Confidence: Automated MRI labels mislead fairness audits in spine segmentation

A new study by Linus Juni, Aasa Feragen, and Aditya Parikh (arXiv 2607.07852) reveals a critical flaw in fairness audits for medical image segmentation: using automated (silver) labels instead of expert (gold) labels can dramatically distort both performance and fairness metrics. The team conducted the first fairness audit of cervical spine MRI segmentation across sex, age, and race using the CSpineSeg dataset. While the deployed model itself appeared demographically fair, the choice of reference label was not neutral. Because silver labels are generated by a model trained on the same gold labels, any new model trained on those gold labels agrees more with the silver labels than with expert truth.

Scoring identical predictions against silver rather than gold overestimates performance by roughly 8 Dice points. More troublingly, it turned the fairness verdict for age from non-significant to significant — not by inflating the gap (false magnitude) but by collapsing within-group variance (false confidence). The authors argue that reference-label provenance is a first-order confounder: performance and fairness should always be reported against expert labels, and any fairness claim must state the provenance of its reference. This work has immediate implications for any fairness audit in medical AI that relies on large, cheaply labeled datasets.