Compartment-Resolved Imaging Traits as a Fix for Ancestry Bias in Opportunistic Bone and Retinal Screening

Medical images are increasingly mined for numbers they were never ordered to produce. A routine abdominal CT scan can yield a bone density reading, and a retinal angiogram can yield a vessel density reading. Most of these numbers are aggregates: one value averaged over tissue compartments that behave differently. A whole-vertebra density mixes the soft inner trabecular bone, where osteoporosis starts, with the dense outer cortical shell. A full-depth retinal vessel map mixes the large superficial vessels with the fragile deep capillaries, where early disease shows up. This proposal argues that such aggregates carry a hidden ancestry bias. The proportion of each compartment depends on body and organ geometry, and geometry varies across populations. A single cutoff on an aggregate can therefore have different sensitivity in different groups even when the underlying biology is identical. Measuring each compartment separately should remove most of this bias.

The test bed is automated vertebral segmentation, meaning deep networks that find, label, and outline each vertebra in a CT scan. A second, smaller replication uses depth-resolved optical coherence tomography angiography (OCTA), a dye-free laser imaging method that resolves retinal blood flow layer by layer. Ancestry enters at two levels. The first is diagnostic thresholds. The second, as a conditional final step, is genetic instruments used in Mendelian randomization (MR), a method that uses inherited variants as natural experiments to estimate cause and effect. The deliverable is an ancestry-stratified audit that splits each pipeline's fairness gap into its sources: segmentation error, vertebral-level mislabeling, compartment mixing, and residual biology. The audit is paired with a test of whether compartment-resolved traits close the gap better than simply adjusting aggregates for body size.

Core Claim

Write an aggregate imaging trait as a volume-weighted mix of compartments:

$$A = w\,T + (1-w)\,C,$$

where $T$ is the compartment that tracks disease (trabecular density, or deep-plexus vessel density), $C$ is the other compartment (cortical shell, or superficial plexus), and $w$ is the fraction of the measured volume that belongs to $T$.

Suppose fracture risk depends only on $T$ and the relation is the same in every population. If the distribution of $w$ differs between populations, for example through vertebral size and shape, then a fixed threshold on $A$ classifies people with the same $T$ differently by group.

The claim is that a substantial part of the ancestry-dependent miscalibration of opportunistic bone screening comes from this compartment mixing, plus mislabeled vertebral levels. Thresholds on trabecular density alone should show smaller between-group sensitivity gaps at matched specificity than either whole-vertebra density or geometry-adjusted aggregates. What is new is not the existence of ancestry gaps, which Precision diagnosis for monogenic diabetes requires ethnicity specific criteria for genetic testing already documents for diabetes. It is a mechanistic decomposition of those gaps in imaging-derived traits, and a test of measurement design as the remedy instead of separate thresholds per group.

Why This Gap Is Real

The vertebral literature supplies the method and shows that robustness has been tested only along physical axes. Automated Opportunistic Trabecular Volumetric Bone Mineral Density Extraction Outperforms Manual Measurements for the Prediction of Vertebral Fractures in Routine CT and Incidental vertebral fracture prediction using neuronal network-based automatic spine segmentation and volumetric bone mineral density extraction from routine clinical CT scans show that automatically isolated trabecular density predicts fractures better than hand-placed regions of interest. Neither stratifies by population. Automated extraction of vertebral bone mineral density from imaging with various scan parameters: a cadaver study with correlation to quantitative computed tomography tests the SpineQ pipeline across tube voltages and slice thicknesses on two cadavers.

That cadaver study also motivates the proposal from an unexpected direction. Its error grew more from one vertebra with heavy osteophytes (bony overgrowths) than from any scanner setting. This suggests that patient anatomy, not acquisition, is the weak point, and anatomy is what plausibly varies with ancestry. AI-based CT assessment of 3117 vertebrae reveals significant sex-specific vertebral height differences shows that vertebral geometry differs enough between groups to require separate reference curves, but it does this for sex, not ancestry.

VerSe: A Vertebrae Labelling and Segmentation Benchmark for Multi-detector CT Images documents the "one-label shift": a network outlines the spine well but names every vertebra one level off. The benchmark also shows that models trained on its less anomalous 2019 release struggled with the more anomalous 2020 cases. Transitional vertebrae, such as an extra lumbar segment or a fused L5, are exactly the kind of anatomy whose frequency might differ by population. Because density varies by spinal level, a label shift silently changes which bone is measured.

On the retinal side, ROSE: A Retinal OCT-Angiography Vessel Segmentation Dataset and New Model and OCTA-500: A retinal dataset for optical coherence tomography angiography study supply layer-resolved segmentation models and data. Tao et al. and Romahn et al. show that superficial and deep layers track different brain pathologies. Together these papers establish that the retinal layers are distinct compartments. None of them asks whether layer composition, or the aggregate's meaning, shifts across populations, and the research file itself flags the absence of multi-ethnic OCTA normal baselines.

On ancestry and causal inference, Causal associations between risk factors and common diseases inferred from GWAS summary data supplies the GSMR method, which uses many genetic variants at once and filters out variants that behave inconsistently (its HEIDI-outlier step). The authors note that its estimates are European-bound. Mendelian randomization study shows no causal relationship between psychiatric disorders and glaucoma in European and East Asian populations is the closest cross-ancestry template, replicating MR in UK Biobank/FinnGen and China Kadoorie Biobank/BioBank Japan. It uses only coarse diagnosis codes, and no imaging traits.

Research Program

The first component builds one opportunistic pipeline, run identically in all populations. It uses nnU-Net-style vertebral segmentation (a self-configuring segmentation network) plus level labeling, then splits each vertebral body into trabecular core and cortical shell. From each scan it outputs whole-body density, trabecular density, cortical attenuation, vertebral height and cross-sectional geometry, and the trabecular fraction $w$.

The cohorts are retrospective clinical CT populations with recorded self-identified race or ethnicity and, wherever linked biobank genotypes exist, genetically inferred ancestry, together with incident fracture follow-up. The research file names no such multi-ancestry CT dataset. Assembling it, from multi-site hospital archives and from the East Asian and European biobanks named in the file where imaging is linked, is the main practical risk and the first milestone.

The second component is the decomposition. For each ancestry group the analysis measures four things against a radiologist-adjudicated subset:

  1. Segmentation quality: Dice overlap.
  2. Level labeling: identification rate.
  3. Transitional anatomy: the prevalence of transitional vertebrae.
  4. Compartment mixing: the distribution of $w$.

The between-group difference in fracture-detection sensitivity at a common threshold is then attributed stepwise. Ground-truth levels are substituted first, then ground-truth masks, then trabecular-only measurement, then geometry adjustment. What remains after all four is called residual biology.

The third component is the retinal replication. Full-depth versus superficial-only versus deep-only vessel density are compared using IPN-V2 and OCTA-Net models in cohorts where ancestry is recorded. The question is whether the gap in a downstream label (for example cognitive impairment, following Hao et al.) shrinks for layer-resolved traits. This is deliberately secondary: the OCTA cohorts in the file are small, and their ancestry composition is not reported.

The fourth component is conditional on genotyped imaging subsets existing in at least two ancestries. It runs association scans for trabecular versus whole-vertebra density, then cross-ancestry MR with ancestry-matched linkage-disequilibrium panels, meaning reference data on which nearby variants are inherited together in each population. This step addresses the cross-ancestry mismatch the file highlights. The test is whether variants instrumenting the aggregate trait show more cross-ancestry heterogeneity and more pleiotropic outliers flagged by HEIDI than variants instrumenting the trabecular trait. The aggregate's variants should partly instrument geometry, which would make them less transportable. Bidirectional MR and Steiger directionality tests will be run, given the feedback loops reported for osteoporosis biomarkers in the PMOP metabolite MR study.

Baselines

The naive reference point is the standard practice of applying one published opportunistic density cutoff to whole-vertebra or mean-attenuation measurements, unchanged across populations, at the default L1–L2 levels. This is what clinical deployments do today, and every other comparison is measured against its between-group sensitivity gap.

The strongest existing method is automated trabecular density extraction as in the Automated Opportunistic Trabecular vBMD paper and the SpineQ pipeline. On reference QCT (CT calibrated against a physical density phantom), SpineQ reports intraclass correlation 0.992 and root-mean-square error 8.93 mg/cm³ (12.65 mg/cm³ with the osteophyte vertebra included), per Ramschütz et al.. For segmentation and labeling, the strongest reported numbers are the VerSe 2020 winner's 96.6% identification rate and 91.72% Dice, per VerSe. No reported result was found on Lacuna for ancestry-stratified performance of any of these. Establishing that number is itself a contribution. The trabecular pipeline will be run with identical weights, preprocessing, and calibration across groups.

Beating it does not mean better overall accuracy. It means a smaller between-group difference in sensitivity at matched specificity for incident vertebral fracture, with no loss in pooled area under the curve. The main ablation adds only geometry adjustment: whole-vertebra density regressed on vertebral height and cross-section, in the spirit of the sex-specific nomograms of the 3117-vertebrae study, with the same tuning budget. This ablation isolates whether compartment separation, rather than size correction, is the active ingredient. A second ablation fixes level labels to ground truth, isolating label-shift effects.

For the retinal arm, the baselines are OCTA-Net, which reports AUC 0.9453 on ROSE-1 superficial-layer segmentation, and IPN-V2, which OCTA-500 reports as about 10% mIoU above the original IPN. Both will be evaluated stratified by group. For MR, the baseline is GSMR run with European reference panels on the aggregate trait.

Why not just…?

Why not just audit existing pipelines stratified by ancestry and stop there? An audit can show that a gap exists but not why. Without the stepwise substitution of true labels, true masks, and compartment-only measurement, one cannot say whether the remedy is better segmentation, better labeling, or a different trait. The audit is the first experiment of this proposal, not a substitute for it.

Why not just adjust aggregate density for body or vertebral size? This is the real threat. If $w$ is almost entirely determined by measurable geometry, then a covariate adjustment recovers what compartment separation recovers, more cheaply. The geometry-adjusted ablation settles this directly. If it closes the gap as well as trabecular-only measurement, the claim fails as stated. The practical output then becomes ancestry-agnostic, geometry-adjusted reference curves, and the proposal reports that outcome rather than arguing around it.

Why not just use ancestry-specific thresholds, as the monogenic diabetes work recommends? That study shows separate criteria can fix miscalibration. But separate thresholds need labeled outcome data in every group, use ancestry labels as a stand-in for mechanism, and fail for admixed individuals. Ancestry-specific thresholds will be included as an upper reference: the experiment reports how much of their gain a single compartment-resolved threshold recovers.

Why not just fine-tune the segmentation network on more diverse scans? The cadaver evidence suggests segmentation is not the bottleneck: error tracked anatomy more than acquisition. The decomposition will show directly what fraction of the gap disappears when ground-truth masks replace predicted ones. If that fraction is large, fine-tuning is the right fix and the plan will pivot to it.

Proposed Experiments and Decisive Checks

Experiment one establishes per-ancestry identification rate, Dice, transitional-vertebra prevalence, and $w$ distributions on an adjudicated subset of several hundred scans per group. The labeling-error hypothesis is supported if identification rate falls in groups with more transitional anatomy, and if, among scans that are labeled one level off, measured density shifts in a consistent direction.

Experiment two is the decisive one. On the full cohort with fracture follow-up, it compares the between-group sensitivity gap at matched specificity across six measurement designs:

  • whole-vertebra density;
  • geometry-adjusted whole-vertebra density;
  • trabecular density;
  • trabecular density with ground-truth levels;
  • trabecular density with ground-truth masks;
  • ancestry-specific thresholds.

Confidence intervals come from bootstrap resampling over patients. The claim survives only if trabecular density narrows the gap significantly more than geometry adjustment does, while pooled discrimination stays within noise.

Experiment three repeats the same comparison of full-depth versus superficial-only versus deep-only density in the OCTA cohorts and is treated as exploratory, given their sample sizes.

Experiment four, run only if genotyped imaging subsets exist in two or more ancestries, compares cross-ancestry effect-size concordance and the HEIDI outlier rate for aggregate- versus trabecular-trait instruments. Compartment-resolved traits should show fewer outliers and closer agreement if mixing with geometry is the source of non-transportability. A null result here would separate the diagnostic claim from the genetic one, which is informative in itself.

References

  1. Precision diagnosis for monogenic diabetes requires ethnicity specific criteria for genetic testing
  2. Automated Opportunistic Trabecular Volumetric Bone Mineral Density Extraction Outperforms Manual Measurements for the Prediction of Vertebral Fractures in Routine CT
  3. Incidental vertebral fracture prediction using neuronal network-based automatic spine segmentation and volumetric bone mineral density extraction from routine clinical CT scans
  4. Automated extraction of vertebral bone mineral density from imaging with various scan parameters: a cadaver study with correlation to quantitative computed tomography
  5. AI-based CT assessment of 3117 vertebrae reveals significant sex-specific vertebral height differences
  6. VerSe: A Vertebrae labelling and segmentation benchmark for multi-detector CT images
  7. ROSE: A Retinal OCT-Angiography Vessel Segmentation Dataset and New Model
  8. OCTA-500: A retinal dataset for optical coherence tomography angiography study
  9. Retinal microvasculature and imaging markers of brain frailty in normal aging adults
  10. Association of retinal vessel pathology and brain atrophy in relapsing-remitting multiple sclerosis
  11. Causal associations between risk factors and common diseases inferred from GWAS summary data
  12. Mendelian randomization study shows no causal relationship between psychiatric disorders and glaucoma in European and East Asian populations
  13. Early detection of dementia through retinal imaging and trustworthy AI
  14. A Mendelian randomization study investigating the causal relationship between 35 blood and urine metabolite biomarkers and postmenopausal osteoporosisShort Communication