Who Does Your Algorithm Fail? Investigating Age and Ethnic Bias in the MAMA-MIA Dataset

AI summary

Overview

• Research area: Fairness and bias auditing in medical image segmentation (computer vision applied to breast cancer MRI), with a focus on demographic disparities in automated tumor segmentation.
• Technical level: Intermediate. The paper uses standard segmentation metrics and well-known statistical tests, but assumes familiarity with fairness auditing concepts such as demographic parity and disparate impact.
• Scope: A fairness audit of the automated segmentation labels released as part of the MAMA-MIA breast cancer DCE-MRI dataset (1,506 T1-weighted cases), evaluating segmentation quality across age groups, ethnic groups, and data sources.

What This Paper Is About

Deep learning models are increasingly used to outline (segment) breast tumors in MRI scans, but most fairness work in medical imaging focuses on classification rather than segmentation, and few datasets combine imaging with demographic metadata. This paper audits the automated segmentation masks shipped with the MAMA-MIA dataset, asking whether those masks are systematically worse for certain patients based on age or ethnicity. The goal is to determine whether observed disparities are intrinsic to the model, artifacts of which hospital or cohort the data came from, or simply the result of some groups being underrepresented.

Key Contributions

1. A first comprehensive fairness audit of MAMA-MIA combining multiple sensitive attributes. The authors state this is, to their knowledge, the first audit of the MAMA-MIA breast tumor segmentation dataset that examines the intersection of multiple sensitive attributes, revealing statistically significant age- and ethnicity-based performance disparities.
2. Evidence that multi-center aggregation masks site-specific ethnic bias. By comparing global analysis to per-source analysis, they show that pooling data from multiple centers obscures ethnic performance gaps that are visible within individual cohorts.
3. A controlled experiment testing whether bias comes from underrepresentation. Training a standard nnU-Net on an age-balanced cohort (n=1,047) isolates representational imbalance as a cause and shows a fairness gap persists without it.
4. Metric-dependent conclusions about bias type. The paper shows that whether ethnic disparity looks intrinsic or driven by data-source confounding depends on whether volumetric overlap (Dice) or boundary accuracy (HD95) is used.

Main Findings

• Age bias favors older patients. Across the full cohort, segmentation quality improves with age. An OLS regression of performance on age found a significant but small relationship (Dice: R² = 0.0104, p = 0.0001; HD95: R² = 0.0093, p = 0.0009).
• The age gap is large in fairness terms. For Dice, the Demographic Parity Difference was 0.0887, with the "young" group achieving high performance at only 70% the rate of the "older" group (Disparate Impact Ratio = 0.699).
• Data source is a significant factor but does not explain the age effect. An ANOVA comparing a baseline model to a source-adjusted model (Performance ~ AgeGroup + DataSource) confirmed the source dataset mattered (Dice: F = 11.76, p = 1.3 × 10⁻⁷; HD95: F = 11.07, p = 3.4 × 10⁻⁷), yet the age effect remained highly significant after adjustment.
• Age bias magnitude varies by data source. An interaction effect between age and dataset was found for Dice (p = 1.6 × 10⁻⁸).
• Balanced training does not remove the age gap. On the age-balanced cohort trained only on expert gold labels, a statistically significant fairness gap of 0.0399 remained (ANOVA p = 0.0260), supporting the conclusion that the bias is intrinsic.
• Table 1 age-stratified results. Balanced cohort Dice scores: Young 0.7304 ± 0.2333, Middle 0.7333 ± 0.2253, Older 0.7703 ± 0.1899 (fairness gap 0.0399, p = 0.0260). Automated labels: Young 0.8082 ± 0.2193, Middle 0.8204 ± 0.2139, Older 0.8612 ± 0.1679 (fairness gap 0.0530, p = 0.0006).
• Global ethnic analysis is contradictory across metrics. For Dice, a non-significant Kruskal-Wallis test (H = 5.09, p = 0.166) and a near-equitable DIR of 0.89 suggested minimal disparity. For HD95, the analysis indicated a significant disparity against the Asian subgroup (p = 0.0046) with a DIR of 0.52.
• Data source strongly confounds ethnic results. ANOVA (Performance ~ Ethnicity + DataSource) found data source highly significant for both Dice (F = 11.78, p = 1.2 × 10⁻⁷) and HD95 (F = 9.10, p = 6.0 × 10⁻⁶).
• Aggregation hides a site-specific gap. The global DPD in Dice scores was only 3.0%, but within the ISPY2 cohort it amplified to 10.0% — a disparity entirely masked by pooling the data.
• The type of bias depends on the metric. Adjusting for data source reduced the Dice effect size by only 6.2% (suggesting intrinsic bias), but reduced the HD95 effect size by 64.0% (suggesting confounding). The authors interpret this as the model producing different types of segmentation errors for certain ethnic groups.

Methodology in Plain English

The dataset provides two sets of tumor outlines for the same MRI cases: "gold" labels drawn manually by a panel of 16 expert radiologists, and "silver" labels produced by an automated model trained on external data. The experts also gave each silver label a qualitative rating (Good, Acceptable, Poor, or Missed).

To measure quality, the researchers compared silver labels against gold labels using two standard segmentation metrics — Dice score (how well the regions overlap volumetrically) and HD95, or 95th percentile Hausdorff Distance (how far off the region boundary is) — alongside the expert ratings.

They then split the data by age group (Young: under 40, 23.2%; Middle: 40–55, 50.1%; Older: over 55, 26.6%) and by ethnicity (Caucasian 74.9%, African-American 16.0%, Asian 5.7%, other minority groups 3.4%), and measured how performance differed. Fairness was quantified with Demographic Parity Difference (the absolute difference in the rate of good outcomes between two groups) and Disparate Impact Ratio (the ratio of the smaller rate to the larger one), classifying a sample as high-performing if it scored in the top 25% of a metric. They also used a "fairness gap": the difference in mean performance between the best and worst subgroups.

Statistically, they used OLS regression to relate sensitive attributes to performance, and because Shapiro-Wilk tests showed the performance distributions were non-normal, they used the non-parametric Kruskal-Wallis H-test for group differences, with post-hoc pairwise comparisons and Bonferroni correction. Categorical expert ratings were analyzed with the Chi-square test. A key step was refitting models with the data source added as a variable, to test whether apparent bias was really just a difference between hospitals.

Finally, to test whether bias came from underrepresentation, they trained their own nnU-Net with 5-fold cross-validation on an age-balanced cohort of 1,047 cases — built by downsampling the Middle and Older groups to match the Young group (n=349) — using only gold labels.

Why This Matters

Impact on research. Fairness audits in medical imaging have concentrated on classification tasks. This work extends the audit paradigm to segmentation and demonstrates two methodological lessons: that multi-center data aggregation can hide subgroup harm, and that conclusions about whether bias is intrinsic can flip depending on which metric is used.

Real-world applications.

• Clinical deployment of tumor segmentation tools: Identifying that younger patients receive lower-quality automated outlines points to where quality assurance should focus before such tools reach clinics.
• Dataset curation and release practices: The findings argue for reporting per-cohort, not just aggregate, performance so that site-specific disparities are visible to downstream users.
• Regulatory and procurement evaluation: The DPD, DIR, and fairness gap framework offers a concrete way to require subgroup-level reporting for medical AI products.
• Treatment planning and monitoring: Because segmentation feeds into diagnosis, monitoring, and treatment planning, disparities in outline quality can propagate into downstream clinical decisions.

Industry relevance. For developers of medical imaging software, the results suggest that releasing a single dataset-level fairness number is insufficient; models may look equitable in aggregate while failing specific subgroups at specific sites. The paper also notes that unaddressed bias could be "compounded across clinical decision points and amplified through iterative model development."

Future Directions

• Investigate the causal mechanisms underlying the persistent bias against younger patients, which the authors hypothesize may relate to physiological factors such as breast tissue density.
• Run controlled training experiments to pin down where in the pipeline the bias is introduced.
• Systematically examine annotation quality for evidence of label bias, since the dataset's dual labeling structure (gold and silver labels plus expert quality ratings) makes this possible.
• Develop targeted mitigation strategies aimed at equitable model performance, which the authors describe as the ultimate goal of the audit.

Target Audience

Researchers and practitioners working on medical image segmentation and medical AI fairness; dataset creators and maintainers who need to decide how to report subgroup performance; clinicians and regulatory reviewers evaluating whether an automated tool is safe across patient populations; and graduate students looking for a concrete example of how to carry out a fairness audit beyond classification.

Funding: Novo Nordisk Foundation, project number 0087102. The paper is licensed CC BY 4.0 and was prepared for the MedEurIPS workshop.

Authors’ abstract

Deep learning models aim to improve diagnostic workflows, but fairness evaluation remains underexplored beyond classification, e.g., in image segmentation. Unaddressed segmentation bias can lead to disparities in the quality of care for certain populations, potentially compounded across clinical decision points and amplified through iterative model development. Here, we audit the fairness of the automated segmentation labels provided in the breast cancer tumor segmentation dataset MAMA-MIA. We evaluate automated segmentation quality across age, ethnicity, and data source. Our analysis reveals an intrinsic age-related bias against younger patients that continues to persist even after controlling for confounding factors, such as data source. We hypothesize that this bias may be linked to physiological factors, a known challenge for both radiologists and automated systems. Finally, we show how aggregating data from multiple data sources influences site-specific ethnic biases, underscoring the necessity of investigating data at a granular level.