🤖 AI

Demographically-Conditioned Synthetic Medical Images for Bias Mitigation and Bias Detection in Disease Classifiers

This paper proposes using a demographically-conditioned synthetic generator to both mitigate training bias through pretraining and detect evaluation bias in medical image classifiers, demonstrating that synthetic minority cohorts can overcome sample-size limitations to improve classifier efficiency and provide reliable fairness audits where real data is scarce.

Original authors: Mahmoud Ibrahim, Bart Elen, Chang Sun, Gokhan Ertaylan, Michel Dumontier

Published 2026-07-17

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mahmoud Ibrahim, Bart Elen, Chang Sun, Gokhan Ertaylan, Michel Dumontier

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a robot to recognize different types of fruit. You show it thousands of pictures of red apples, but only a handful of green pears. The robot will become a master at spotting apples but might get confused or make mistakes when it sees a pear, simply because it hasn't seen enough of them. This is a common problem in "machine learning," where computers learn by looking at data. In the world of medical imaging, this gets much more serious. If a computer program is trained to spot diseases in CT scans, but the training data mostly comes from young men, the program might fail to spot the disease in elderly women. This isn't just a glitch; it's a form of bias that could lead to unfair or dangerous medical decisions.

To fix this, scientists often try to "audit" the computer, checking how well it works on different groups of people. But here's the catch: in the real world, there are often very few pictures of those minority groups (like elderly women) to test with. It's like trying to judge how good a chef is at cooking vegan dishes when you only have three vegan meals to taste. The sample size is too small to know for sure if the chef is actually bad or if you just got unlucky with the three meals you tried. This paper tackles that exact problem: how do we train fair medical AI when we don't have enough real data, and how do we test it fairly when our test sets are too small?

The researchers behind this study decided to try a clever trick: they built a "digital twin" generator. Think of this as a magical art machine that can create brand-new, fake medical images that look real but are made up by a computer. They taught this machine to create CT scans specifically for the groups that are usually missing, like elderly women or young men. They then used these fake images in two ways. First, they used them to "pre-train" the medical AI, giving it a broad, balanced foundation before showing it the few real images it had. Second, they used the fake images as a massive test set to see if the AI was actually being fair to everyone.

The results were surprisingly effective. When they used the fake images just to add more data to the training pile (mixing them in with real images), it did improve the AI's performance over using real data alone, but it wasn't the best approach. It was like trying to learn a language by reading a dictionary and a novel at the exact same time; the computer got a bit confused about the balance. However, when they used the fake images first to build a strong foundation, and then fine-tuned the AI with just a tiny amount of real data (only 1% of the usual amount), the results were amazing. The AI became incredibly accurate and fair, performing even better than models trained on huge amounts of real data. It was as if the AI learned the "grammar" of the disease from the fake books and then just needed a few real sentences to perfect its accent.

On the testing side, the fake images proved to be a superhero for fairness checks. When the researchers tried to test the AI on a small group of real patients, the results were shaky and unreliable because there were so few people in the test group. But when they tested the AI on a huge group of "fake" patients generated by their machine, the results were much more stable. The fake test set could perfectly predict which groups the AI was struggling with (ranking them correctly), matching the results you would get if you had a massive, perfect real-world test set. In fact, for the groups with the fewest real patients, the fake test set was several times more reliable than the real one. However, it's important to note that while the fake data was great at identifying where the bias was, the exact numbers (like specific accuracy scores) did shift slightly compared to real data, depending on how the computer decided to make its final call.

The paper explicitly argues against the idea that you should just mix fake and real data together in one big pot during training. They found that while this "joint augmentation" approach did help, the "two-stage" approach (learning from fake data first, then refining with real data) was far superior. The secret sauce was the order of operations: learn the basics from the balanced fake data first, then refine with the real data later. They also showed that while fake data is great for spotting where the bias is (ranking the groups), you still need some real data to get the exact numbers right for groups that are well-represented.

In short, this paper suggests that we don't need to wait decades to collect enough real data to make fair medical AI. By using a smart, two-step training method with synthetic images, we can create fair classifiers with a fraction of the real data, and we can use those same synthetic images to reliably audit the system for bias, especially for the people who are usually left out of the conversation. It's a promising step toward making sure medical technology works for everyone, not just the majority.

Technical Summary: Demographically-Conditioned Synthetic Medical Images for Bias Mitigation and Bias Detection

1. Problem Statement

Medical image classifiers face a dual challenge regarding per-subgroup fairness audits:

Training Bias: Training data often under-represents minority subgroups (e.g., specific sex and age combinations), causing classifiers to underperform on these groups.
Evaluation Bias (Sample-Size Problem): Held-out real test sets are dominated by majority demographics. Minority subgroups often have only a few dozen samples per cell. At this scale, confidence intervals for per-subgroup performance are wider than the bias itself, making it impossible to distinguish real performance gaps from sampling noise.

The authors argue that a demographically-conditioned synthetic generator can address both sides of this problem: filling training gaps to mitigate bias and generating large synthetic minority cohorts to enable statistically stable evaluation.

2. Methodology

Dataset and Partitions
The study utilizes the COVIDx-CT-3A dataset (Variant-A subset), comprising 168,694 axial chest CT slices from 1,718 patients. Data is stratified by sex (Male/Female) and age (5 groups), creating 10 subgroups.
The data is split into four patient-disjoint partitions:

Train/Val: Standard training and validation sets.
Hold-out: Matches the joint distribution of train/val; used for generator fine-tuning.
Hold-out-minority: Enriched for minority cells to create a "well-powered" oracle.
Derived Test Sets:
- Real-test-small: A small, realistic audit set (16,761 slices) mimicking deployment scarcity.
- Real-test-large: A well-powered oracle (71,037 slices) created by merging and re-splitting the hold-out and hold-out-minority sets.

Generator Architecture
The authors fine-tune Stable Diffusion 2.1 end-to-end on the hold-out partition.

Conditioning: Natural language prompts encode age, sex, and diagnosis (e.g., "{age} year old {SEX}. {COVID-19 findings/Normal} chest CT scan").
Training: The U-Net denoiser and CLIP text encoder are fine-tuned for 20,000 steps; the VAE remains fixed. Classifier-free guidance is used with 10% null-prompt dropout.
Baseline: A latent diffusion model pre-trained on MIMIC-CXR (chest X-rays) and fine-tuned on CTs serves as the comparison baseline.

Downstream Classifier Regimes
A DenseNet-121 binary classifier is trained under 22 different regimes to test two main hypotheses:

Bias Mitigation (Training):
- Augmentation (AUG): Mixing fixed synthetic balanced data with varying amounts of real data.
- Two-Stage Pretraining (FT): Pretraining on synthetic data (Balanced or Distribution-Matched) followed by fine-tuning on small fractions of real data (1% to 100%).
Bias Detection (Evaluation):
- Evaluating a fixed real-data-trained classifier on three test cohorts: Real-test-large (Oracle), Real-test-small (Real comparator), and Synth-test-large (Synthetic proxy).

3. Key Contributions

1. A Demographically-Conditioned Chest CT Generator
The authors successfully fine-tuned Stable Diffusion 2.1 end-to-end for chest CTs. This generator outperforms a chest X-ray (CXR) pre-trained baseline by 35% on FID-Inception and 32% on FID-RadImageNet, while maintaining high conditioning accuracy for sex (≥0.84) and age (≥0.33) and preserving diagnostic signal (COVID-19 AUC > 0.99).

2. Bias Mitigation: Pretraining as a Prior
The study finds that balanced synthetic data is most effective as a pretraining prior, not as joint augmentation.

Efficiency: Pretraining on a balanced synthetic cohort and fine-tuning on just 1% of real data (585 slices) produces a classifier that surpasses the full-real-data baseline on both mean and worst-cell Matthews Correlation Coefficient (MCC).
Schedule vs. Data: A paired comparison shows that sequential pretraining followed by fine-tuning significantly outperforms joint augmentation using the same data volume. The training schedule (representation building via synthetic data, boundary refinement via real data) is the source of the improvement, not merely the data composition.

3. Bias Detection: Synthetic Cohorts as Auditors
The synthetic minority cohort serves as a reliable proxy for detecting subgroup bias where real data is scarce.

Rank Agreement: The synthetic estimator perfectly reproduces the subgroup ranking of the well-powered real oracle (Spearman $\rho = 1.00$ on MCC and Recall), whereas the small real test set fails to do so ( $\rho \approx 0.6$ ).
Calibration: In minority cells with fewer than 1,500 real samples, the synthetic estimator's bias is 2.3x smaller than the real test set's bias. In well-sampled cells, the real test set remains superior. The two estimators are complementary along a sample-size axis.

4. Results

Generator Quality: The SD 2.1 fine-tuned generator achieves an FID-Inception of 50.14 on the hold-out set, compared to 77.04 for the CXR baseline.
Mitigation Performance:
- FT-Balanced @ 1% Real: Achieves Mean MCC of 0.83 and Worst-Cell MCC of 0.52.
- Full-Real Baseline (TRTR-Biased): Achieves Mean MCC of 0.78 and Worst-Cell MCC of 0.17.
- This represents a 100x real-data efficiency gain, where a tiny real fine-tune on a synthetic prior outperforms training on the full real dataset.
Evaluation Performance:
- Aggregate Metrics: Synthetic test sets show minor drift in threshold-sensitive metrics (F1, MCC) compared to real sets (e.g., MCC drift of -0.035), but threshold-free metrics (ROC-AUC) are nearly identical.
- Subgroup Audit: The synthetic proxy achieves perfect rank correlation with the oracle for MCC and Recall, identifying the worst-performing subgroups accurately even when the real test set is too small to detect them.

5. Significance and Claims

The paper claims that demographically-conditioned synthetic data is not just a data augmentation tool but a structural fix for the sample-size problem in fairness auditing.

Mitigation: It allows institutions with limited labeled minority data to train fair classifiers by using synthetic data as a representation prior, achieving high fairness with minimal real data.
Detection: It provides a statistically stable method to audit classifiers for subgroup bias in scenarios where real test sets are too sparse to yield reliable confidence intervals.
Complementarity: The authors emphasize that synthetic and real estimators are complementary; synthetic data excels in under-sampled minority cells, while real data remains the gold standard for well-sampled cells.

The work concludes that synthetic chest CTs are most useful exactly where fairness audits are weakest: simultaneously acting as a fix for training bias and a measure for evaluation bias in minority subgroups.