False Confidence: Automated Labels Confound Fairness Audits in Cervical Spine Segmentation

False Confidence: How Automated Labels Confound Fairness Audits in Spine MRI

When researchers use AI-generated labels to train or test medical models, they might accidentally create a "biased ruler." This can make a model look much more accurate than it really is. It can also trick scientists into thinking a model is unfair when it actually isn't. This happens by hiding the natural variations in the data.

In medical imaging, automated segmentation is becoming a staple of clinical workflows. This is the process of a computer labeling specific anatomical structures, like vertebrae or discs, in an MRI. However, obtaining "gold labels" (perfect annotations created by human experts) is expensive and slow. To solve this, many modern datasets supplement a small amount of expert data with "silver labels." These are machine-generated approximations.

A new study from researchers at the Technical University of Denmark explores a dangerous side effect of this practice. By conducting the first demographic fairness audit of cervical-spine MRI segmentation, the authors reveal that the choice of reference label is not a neutral act. Depending on whether you use an expert or a machine to grade a model, you might arrive at two different conclusions. You could find that a model treats different ages, races, or sexes unfairly, or you might find it perfectly fair.

The hidden cost of silver labels

Current segmentation workflows rely heavily on hybrid datasets to manage annotation costs. While a handful of scans are meticulously labeled by radiologists, the bulk of the training data often consists of silver labels. These are generated by previous iterations of AI models. The assumption is that these labels are "good enough" to guide learning.

However, the authors argue that this creates a fundamental problem for auditing. An audit is only as good as its ruler. If the ruler (the reference label) is itself biased or shares the same flaws as the model being tested, the audit becomes a circular exercise. In the cervical spine, anatomical structures vary significantly by age and sex. Using a biased ruler doesn't just misrepresent accuracy. It actively obscures or manufactures social disparities.

A mechanism of false confidence

To understand how this happens, we must look at the relationship between the model and the silver labels. The authors identify a phenomenon they call "false confidence." This occurs because of "label leakage." Since silver labels are typically generated by a model trained on the same expert gold labels as the model under audit, the two models share systematic errors.

The mechanism works in three stages: 1. Correlation: The model being audited ($M_{mix}$) and the silver ruler ($M_{gold}$) are both trained on overlapping subsets of expert data. They learn to "see" the anatomy in similar ways. 2. Agreement: When $M_{mix}$ makes a mistake, the silver ruler is likely to make that same mistake. Because they agree on the error, the mathematical score remains high. This score is often the Dice coefficient (a metric measuring the overlap between two shapes). 3. Variance Collapse: Because the models are so similar, the perceived differences between subgroups appear to vanish. The "noise"—the natural variation in how a model performs across different patients—is suppressed.

As seen in, the gold ruler shows a wider spread of performance across age bins.

This reflects the true difficulty of segmenting older patients with degenerative anatomy. In contrast, the silver ruler's boxes are "razor-thin." This means the variance has collapsed. This makes a slight, clinically negligible trend look like a statistically certain pattern.

Inflated accuracy and manufactured significance

The empirical evidence provided by the authors is stark. Testing the $M_{mix}$ model on a set of 76 gold-standard images, the researchers found that scoring the model against silver labels overestimated its performance. It inflated the macro Dice score by approximately 8 points (0.973 vs. 0.897).

More critically, the choice of ruler flipped the fairness verdict for age. When measured against expert gold labels, the model passed all fairness tests. There were no significant disparities. However, when measured against the silver ruler, the authors report that 11 out of 63 statistical tests became FDR-significant (a method used to reduce false positives in multiple testing). Specifically, the silver ruler turned a non-significant age trend into a significant one.

The authors emphasize that this isn't "false magnitude," where a gap is exaggerated. Instead, it is "false confidence." The silver ruler doesn't necessarily make the gap between groups look larger. Instead, it makes the results look more precise than they actually are. This leads researchers to believe they have discovered a bias that isn't truly there.

Limits of the audit

While the findings are a significant warning, the study has notable constraints. The researchers conducted their audit on a single institution's dataset. They used a specific architecture (nnU-Net) and a single MRI sequence. It is not yet clear if this "false confidence" effect persists across different types of medical imaging.

Furthermore, the gold test set was relatively small (76 cases). It included only about 20 Black patients. The authors admit this limits the statistical power of their race-based analysis. Finally, the silver ruler was a reconstruction of the dataset's original generation process. The study demonstrates the effect of a correlated ruler. It does not explore how an entirely independent but biased ruler might behave.

The verdict: report the provenance

If you are building or auditing medical AI, the verdict is clear. You cannot treat machine-generated labels as neutral benchmarks. The authors argue that reference-label provenance is a "first-order confounder" in segmentation evaluation.

For practitioners, the advice is actionable. Always report performance and fairness against expert gold labels whenever possible. If you must use silver labels, you must explicitly state their origin. Furthermore, if your model's performance is significantly higher and its variance is much lower when measured against a silver ruler, be careful. Specifically, if the variance collapses by about 4 times, your benchmark is likely too close to your model to be considered independent. Accuracy without transparency is just false confidence.