Auditing 3D Imaging Models by Tissue, Not by Demographics

Medical imaging models are usually checked for fairness by splitting the test set by age, sex or ethnicity and comparing accuracy across groups. In practice the scans that arrive at a deployed model often carry no such labels, because privacy rules strip them, archives never recorded them, or they sit unlinked in a different hospital system. There is also growing evidence that demographic groups are the wrong unit of analysis anyway. In 3D breast MRI, a fairness gap between young and older women persisted even after the training data was balanced by age, because the real cause was denser breast tissue. In abdominal CT, male-trained models generalized better to women than female-trained models did, because the male cohort had more varied organ shapes. We propose to audit deployed 3D segmentation models by searching for groups of scans that share measurable tissue properties and on which the model fails. The tissue properties are computed from the scan itself: organ shape and size, plus "sub-visual" texture statistics, meaning numerical descriptions of voxel intensity patterns that radiologists cannot see by eye. Per-scan accuracy is estimated without expert masks. The audit needs no demographic labels and no ground truth, and every group it reports comes with a plain description such as "small tumors in dense, low-heterogeneity parenchyma."

The bet is falsifiable. If failures really follow tissue rather than census categories, then the worst group found in tissue-descriptor space should have a larger and more reproducible accuracy gap than the worst demographic group, and it should overlap with the demographic groups that are known to suffer. If the bet is wrong, plain demographic stratification, or stratification by age and sex predicted from the image, will match it. Either outcome tells hospitals what to monitor when the demographic fields are empty.

Core Claim

For volumetric segmentation models, the worst-performing slices of a patient population can be found and described using a compact set of shape and texture descriptors computed from the model's own predicted volume, scored with a mask-free per-scan quality estimate. Here a "slice" means a subset of patients defined by a short, human-readable rule over those descriptors. We claim three things. First, these slices expose larger held-out performance gaps than demographic subgroups do. Second, texture descriptors add detection power beyond organ volume and shape alone, the macro-morphology that Quantifying the Impact of Population Shift Across Age and Sex for Abdominal Organ Segmentation measured. Third, the procedure works as a demographic-free monitor during a simulated silent deployment to a new site.

The idea differs from the closest prior work along specific axes. Unsupervised bias discovery for segmentation, Unsupervised bias discovery in medical image segmentation, estimates Dice without target masks using reverse classification accuracy, registering predictions back to a labeled atlas. It reaches 90% and 92% sign agreement on lung and heart gaps, but it is 2D and still groups results by known sex labels. We reuse its mask-free estimation idea and remove the dependence on labels for grouping. Demographic-free training methods change the model rather than audit it, and they constrain our design. The adversarial reweighting in Fairness without Demographics through Adversarially Reweighted Learning breaks down under label noise, which is endemic in 3D masks. AttEN in Achieve Fairness without Demographics for Dermatological Disease Diagnosis relies on a lesion-versus-background-skin split that has no analogue inside an organ volume. Radiomics work such as Hybrid transformer convolutional neural network-based radiomics models for osteoporosis screening in routine CT uses texture as a predictor and never asks whether it encodes the populations a model fails on. We use texture as the coordinate system for an audit instead.

Research Program

The motivating evidence comes from three audits that all needed curated labels. Who Does Your Algorithm Fail? Investigating Age and Ethnic Bias in the MAMA-MIA Dataset found that tumor segmentation degrades for women under 40 and that the ethnic gap is 3.0% pooled but 10.0% inside the I-SPY2 cohort. Cevora et al. found population shift costs 1–6% Dice on abdominal organs, with a 6% kidney drop from male to female. Analysing race and sex bias in brain age prediction showed that 3D network features separate racial groups. Each of these studies points to tissue or morphology as the mechanism, and none tests it. The program has four components.

The first component computes descriptors. For each scan, we take the target model's predicted mask, for example an nnU-Net tumor or kidney mask. Within that mask and a fixed-width shell around it, we compute about forty pre-registered descriptors. These cover volume, sphericity and surface-to-volume ratio, first-order intensity statistics, and a small fixed set of 3D co-occurrence and run-length texture features. We deliberately avoid the 1,767-feature extraction of Liu et al., because a large descriptor space would make the slice search a false-discovery engine. Intensities are normalized per scan. We also record acquisition tags (scanner model, field strength, slice thickness) separately, since these usually survive de-identification even when demographics do not.

The second component estimates per-scan quality without masks. The primary estimator is disagreement among an ensemble of five independently trained segmentation models, measured as mean pairwise Dice. As a secondary estimator we run a 3D atlas-registration variant of reverse classification accuracy on a subset. A known risk is that the estimator is itself biased across groups, for instance overconfident on dense tissue. We therefore measure the correlation between estimated and true quality separately within every demographic and descriptor stratum.

The third component searches for slices. We fit shallow regression trees of depth at most three, with a minimum leaf size of 5% of the cohort, that predict estimated quality from the descriptors. The lowest-scoring leaf, stated as its rule, is the reported slice. We search on one half of the audit set and report the true Dice gap of the same rule on the other half, so search optimism does not inflate the claim.

The fourth component is monitoring. In a simulated silent deployment, the model trains on some sites and new-site scans stream in. The monitor re-runs the search on rolling windows and raises an alarm when a slice's estimated gap exceeds a threshold calibrated on the training sites. This targets a specific limitation: the dashboard in DEPLOYR: A technical framework for deploying custom real-time machine learning models into the electronic medical record monitors fairness only when demographic fields stream from the record, and the system does not handle imaging at all. We do not build EMR integration. We build the component such a system lacks and evaluate it offline under deployment-like splits.

Baselines

The naive reference point is aggregate monitoring. It reports mean estimated Dice over all incoming scans and raises an alarm on a global drop. It is what a deployment dashboard without demographics would show, and Parikh et al.'s 3.0% pooled versus 10.0% within-site ethnic gap demonstrates how it hides local failure. To beat it, a method must detect a held-out subgroup gap that the global mean misses.

The oracle reference is demographic stratification with true labels, run on the same model and test split. It uses age bins and ethnicity on MAMA-MIA, and the under-50 versus over-70 age split and sex on TotalSegmentator/AMOS, following Cevora et al. The relevant reported figures are Parikh et al.'s disparate impact ratio of 0.699 for young versus older patients and their persistent gap of 0.0399 after age balancing, and Cevora et al.'s 6% kidney Dice drop from male to female. Our first success criterion is that the tissue slice, found without labels, shows a held-out gap at least as large as the worst demographic group's gap, using the same minimum group size.

The strongest label-free competitor is slicing in the model's own representation space. We take bottleneck features of the same nnU-Net, cluster them with k-means over a tuned number of clusters, and report the worst cluster by estimated quality. We run the same search over embeddings from the SegVol 3D foundation model (SegVol: Universal and Interactive Volumetric Medical Image Segmentation). This baseline is the strongest because Piçarra et al. showed that deep 3D features already encode demographic structure, so embedding clusters may find the same failures with no hand-designed descriptors. For fairness, both arms use the identical quality estimator, the same search/confirm split and minimum slice size, and an equal number of hyperparameter configurations. No reported result was found on Lacuna for embedding-space slice discovery on 3D segmentation. Beating it means a larger held-out gap, or an equal gap with a rule that radiologists in a small reader check can restate correctly.

Two further baselines address cheaper sources of the same information. The predicted-demographics baseline trains a 3D regressor for age and a classifier for sex on training scans, then stratifies deployment scans by these predictions. The acquisition-tag baseline runs the same tree search over scanner, site and slice-thickness tags only. The key ablation repeats our method with shape and volume descriptors only, dropping texture. If texture adds nothing, the sub-visual component of the proposal should be cut, and the method collapses to Cevora et al.'s morphological-diversity idea made label-free.

Why not just…?

Why not just predict age and sex from the scan and stratify by that? This is the most serious threat for age and sex, which are readily recoverable from CT and MRI. It does not help with ethnicity, which is weakly and unreliably encoded. More importantly, the evidence says demographic groups are imperfect proxies for the failure mechanism: Parikh et al.'s gap survived age balancing, and Cevora et al.'s generalization ran opposite to naive demographic matching. The predicted-demographics baseline settles the question directly. If it matches the tissue slices' held-out gaps on both datasets, we will report that predicted demographics suffice for this purpose.

Why not just cluster the network's embeddings? Embeddings may well capture the same structure, and they are a real competitor. Their drawback is that a failing cluster has no description: a clinician cannot act on "cluster 7," and Henry et al. found that trust comes from experience rather than opaque explanations. Readable rules matter for whether a monitor is used at all. Still, if embedding clusters find larger gaps, the accurate conclusion is that descriptors trade detection power for legibility. Our experiments measure that trade-off rather than assume it away.

Why not just slice by scanner and site, which are usually recorded? Acquisition is a likely confounder, since texture features are notoriously scanner-sensitive. A tissue slice could simply be "scanner X" in disguise. The acquisition-tag baseline, together with a check of how much of each discovered slice's membership is explained by tags, distinguishes biological slices from acquisition slices. If acquisition explains everything, the cheap tag-based monitor wins.

Why not just collect demographics? Legacy archives such as Cam-CAN cannot be backfilled (The diversity and disparity in biomedical informatics (DDBI) workshop (PSB 2018)). Even complete labels miss tissue-driven gaps like the one that survived age balancing on MAMA-MIA. We treat collected demographics as the oracle to compare against, not as a substitute for the audit.

Proposed Experiments

The primary testbed is MAMA-MIA: 1,506 DCE-MRI scans from DUKE, I-SPY1, I-SPY2 and NACT, with age and ethnicity labels. We train nnU-Net ensembles on three cohorts and treat the held-out cohort as the deployment site, rotating through all four. The second testbed is TotalSegmentator (1,082 CT scans) with AMOS (500) as the shifted site, using the age and sex labels from Cevora et al. and AMOS's own metadata, with kidneys and liver as targets. Labels and true masks are used only to score the audit, never inside it.

The first decisive check compares held-out true Dice gaps of the worst slice across our method, the embedding baseline, the predicted-demographics baseline, the tag baseline and the true-demographics oracle. We report 95% bootstrap intervals and use a matched minimum group size. The second check measures recovery: the overlap, as Jaccard index, between our worst slice and the under-40 group on MAMA-MIA and the female group for kidneys on TotalSegmentator. We expect partial overlap, plus extra patients who are tissue-similar but outside the demographic group and who also fail. The third check is the texture ablation. The fourth check validates the quality estimator per stratum. If the ensemble-disagreement estimator understates errors in the very slices we care about, the audit is invalid there, and we will report it as such.

For monitoring, we inject controlled shifts into the deployment stream by enriching it with young patients or I-SPY2 scans at increasing rates. We measure detection delay in number of scans and false-alarm rate on unshifted streams. For context, Deep Hypothesis Tests Detect Clinically Relevant Subgroup Shifts in Medical Images reported that classifier two-sample tests detect subgroup shifts with 100–200 images on 2D data, so we include such a test on embeddings as a shift-detection reference. That test detects that the population changed. Our monitor is meant to report who is now being failed.

The project is worth pursuing further if tissue slices match or exceed the demographic oracle's held-out gap on both datasets and texture adds measurably over shape. It should be abandoned in favor of the cheaper alternative if predicted demographics or acquisition tags explain the same failures.