💻 computer science

FairGen: Preference-Aligned Diffusion for Demographically Equitable Medical Image Synthesis

FairGen is a fairness-aware diffusion framework that synthesizes demographically balanced medical images by embedding physician-aligned preferences, significantly improving subgroup coverage and reducing diagnostic bias across dermatology, radiology, and neuroimaging tasks while maintaining high diagnostic accuracy.

Original authors: Zhimin Li, Ruichen Zhang, Zhen Tan, Howard J Aizenstein, Jingtong Hu, Tianlong Chen

Published 2026-06-16

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Zhimin Li, Ruichen Zhang, Zhen Tan, Howard J Aizenstein, Jingtong Hu, Tianlong Chen

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 computer to recognize different types of skin rashes, lung issues, or brain conditions. To do this, you show it thousands of photos. But here's the problem: the photo albums we have are very unbalanced. They are full of pictures of light-skinned people, young adults, and men, but they are almost empty of pictures of dark-skinned people, the elderly, or women.

Because the computer only sees a few examples of these "rare" groups, it gets confused when it sees them later. It might miss a disease in a dark-skinned patient because it was never trained on what that disease looks like on dark skin. This is like teaching someone to drive only on sunny days; when it rains, they don't know what to do.

Enter "FairGen": The Balanced Photo Studio

The researchers in this paper created a tool called FairGen. Think of FairGen as a magical, fairness-focused photo studio. Instead of just copying existing photos, it uses advanced AI to create new, realistic medical images that fill in the missing gaps.

Here is how it works, using simple analogies:

1. The Problem: A Skewed Library

Imagine a library where 90% of the books are about cats, and only 10% are about dogs. If you ask a student to write a report on "pets," they will know everything about cats but will know almost nothing about dogs. In medical imaging, the "books" are patient photos. If a hospital has mostly photos of young, light-skinned men, the AI trained on those photos will be terrible at diagnosing older women or people with darker skin.

2. The Solution: FairGen's Three Magic Tricks

FairGen doesn't just make random pictures; it makes smart pictures to fix the library's imbalance. It uses three specific tricks:

Trick One: The "Equalizer" (Resampling)
Imagine the AI is a chef cooking a stew. If the pot has too many potatoes and not enough carrots, the stew tastes bad. FairGen looks at the ingredients (the data) and says, "We need more carrots!" It specifically targets the missing groups (like dark skin or older brains) and tells the AI to focus extra attention on them while it learns.
Trick Two: The "Variety Guard" (Diversity Loss)
Sometimes, when AI tries to fix a problem, it gets lazy and just copies the same thing over and over. FairGen has a "Variety Guard" that ensures the new photos are all different from each other. It makes sure the AI doesn't just create 100 identical pictures of a dark-skinned patient, but 100 unique and realistic variations.
Trick Three: The "Doctor's Whisper" (Physician-Aligned Preferences)
This is the most important part. Usually, AI just tries to make pictures that look real. But in medicine, looking real isn't enough; the picture must show the right medical details.
The researchers brought in real doctors to act as judges. They showed the AI pairs of generated images and asked, "Which one shows the specific signs of dementia?" or "Which one shows the rash correctly?"
The AI learned from the doctors' choices. It's like a student who doesn't just memorize the textbook but gets a tutor to point out exactly what to look for. This ensures the fake images aren't just pretty; they are medically accurate, even for rare conditions.

3. The Results: A Fairer Future for AI

The researchers tested FairGen on three different types of medical images: skin photos, chest X-rays, and brain scans.

The "Before" Picture: Standard AI models were like a biased judge, often making mistakes for minority groups.
The "After" Picture: When the AI was trained using FairGen's new, balanced photos, it became much fairer.
- For skin images, fairness improved by 95.9%.
- For chest X-rays, it improved by 80.0%.
- For brain scans, it improved by 35.2%.

Crucially, the AI didn't get worse at diagnosing the majority groups. It didn't have to "sacrifice" accuracy for some people to help others; it simply got better at seeing everyone.

What This Paper Actually Says (and Doesn't Say)

What it does: It proves that we can use AI to generate synthetic medical images that fix data imbalances. It shows that these fake images can be used to train better, fairer diagnostic computers.
What it is NOT: The paper is careful to say this is a research tool, not a finished product ready for hospitals yet.
- The "fake" images are for training other computers, not for doctors to look at directly to diagnose a patient.
- The study used existing public datasets and expert reviews to prove the concept works, but it is not a clinical trial where patients were treated.
- The authors warn that while the images look real, they must be used carefully, as AI can sometimes create "hallucinations" (fake details) that look real but aren't true.

In a Nutshell:
FairGen is like a "demographic equalizer" for medical AI. It uses a doctor's guidance to create a balanced set of training photos, ensuring that the next generation of medical AI doesn't just work for the majority, but works well for everyone, regardless of their skin color, age, or gender.

Technical Summary: FairGen

Problem Statement

Medical imaging datasets frequently exhibit severe demographic imbalances due to disparities in healthcare access and differential disease prevalence across populations. These imbalances are compounded by the fact that disease phenotypes can manifest differently across demographic groups (e.g., psoriasis appearing grayish on darker skin rather than salmon pink), rendering certain presentations naturally rare in standard datasets. Consequently, AI models trained on such data risk perpetuating diagnostic bias, leading to significant performance drops for underrepresented subgroups (e.g., up to 50% lower accuracy in Black subpopulations for chest radiography models).

Existing fairness-aware generative approaches face limitations: Generative Adversarial Networks (GANs) often degrade image quality or require curated datasets, while recent diffusion-based attempts rely on costly fine-tuning on balanced datasets or post-hoc controls that may overcorrect or amplify bias. Furthermore, even state-of-the-art diffusion models can produce outputs that are predominantly classified as white and male under neutral prompts. There is a lack of scalable, modality-agnostic frameworks that embed fairness objectives directly into the generation process while preserving pathology-relevant visual features.

Methodology: FairGen Framework

FairGen is a fairness-aware diffusion framework designed to synthesize demographically balanced medical images while maintaining clinical fidelity. The framework operates through a unified pipeline integrating three core components:

Attribute-Disease Resampling:
To address data imbalance at the source, FairGen employs two complementary resampling strategies:
- Class Balanced Random Sampling (CBRS): Assigns weights to underrepresented attribute-disease pairs based on frequency and a gradient-aware mechanism that increases weights for groups with higher training gradients (indicating learning difficulty).
- Square Root Random Sampling (SQRS): Applies a square-root transformation to group frequencies to avoid over-amplification of rare categories, also integrated with gradient-aware adjustments.
Class-Diversity Regularization:
A class-diversity loss ( $L_r$ ) is integrated into the diffusion training objective. This loss penalizes the model when noise predictions across different classes are indistinguishable, enforcing clear distinctions between classes and preventing mode collapse or the dominance of majority groups. It ensures the generation of diverse synthetic samples across all sensitive attributes and disease classes.
Physician-Aligned Preference Guidance:
To ensure generated images preserve clinically meaningful features, FairGen incorporates Direct Preference Optimization (DPO).
- Data Construction: Physicians review pairs of candidate images against specific clinical prompts (e.g., "Age above 75, Demented") and select the image that better matches the target phenotype based on anatomical cues (e.g., widened sulcal spaces) rather than general realism.
- Optimization: A reward model is trained to distinguish preferred from non-preferred samples. This reward signal guides the diffusion model at each timestep to align with physician-endorsed phenotypes, preventing the erasure of subtle pathological features in rare or minority cases.

The framework utilizes a Stable Diffusion v1-4 backbone. During training, the VAE and text encoder remain frozen, while the U-Net is fine-tuned with the added fairness losses. In the downstream stage, the generated synthetic data is combined with real data, and a dynamic, performance-aware reweighting mechanism is applied during classifier training to further reduce residual bias.

Key Contributions

Framework Design: Introduction of FairGen, a modality-agnostic diffusion framework that embeds fairness directly into the generation pipeline via resampling, diversity regularization, and physician-aligned preference guidance.
Physician Alignment: A novel approach to integrating clinical expertise into generative AI by constructing preference datasets where physicians select images based on specific pathological features, ensuring the model learns clinically relevant phenotypes rather than just statistical averages.
Comprehensive Evaluation: A rigorous evaluation across three distinct modalities (dermatology, radiology, and neuroimaging) covering both generative fidelity and downstream diagnostic fairness.

Results

The framework was evaluated on the Fitzpatrick-17k (skin images), CheXpert (chest radiography), and OASIS (T1-weighted MRI) datasets.

Generative Fidelity and Fairness:

Image Quality: FairGen achieved the lowest Fréchet Inception Distance (FID) scores among five baselines (including Stable Diffusion, FairDiffusion, CBRS, SQRS, and CBDM), with relative improvements of 10.1% for skin images, 13.0% for chest radiography, and 3.0% for MRI.
Generative Fairness: FairGen significantly reduced the variance of FID scores across demographic subgroups. For skin images, FID variance dropped from 862.2 (Stable Diffusion) to 317.3, indicating that minority groups are synthesized with fidelity comparable to majority groups.
Visual Evidence: Qualitative analysis and t-SNE projections showed that FairGen produces a homogenously mixed distribution of latent features, whereas baselines exhibited distinct clustering by sensitive attributes.

Downstream Diagnostic Performance:

Fairness Metrics: Models trained with FairGen-generated data showed substantial improvements in fairness metrics (Demographic Parity, Equal Opportunity, Equalized Sensitive Subgroup Performance, and Normalized Accuracy Range).
- Skin Images: 95.9% improvement in fairness.
- Chest Radiography: 80.0% improvement.
- Brain MRI: 35.2% improvement.
Accuracy: The framework maintained competitive diagnostic accuracy relative to models trained on original clinical data, demonstrating that fairness improvements did not come at the cost of performance.
Expert Validation: In blinded pairwise evaluations, experts preferred FairGen-generated images over baselines in 85% of skin image comparisons and 55% of chest radiography comparisons when assessing disease-relevant clinical features.

External Validation:
Performance trends were validated on independent cohorts (ISIC, IXI, MIMIC-CXR), confirming that the framework's benefits extend beyond the original in-distribution datasets, though the authors note these are benchmark-level evaluations.

Significance and Claims

The paper positions FairGen as a research-stage framework for fairness-aware medical image synthesis and subgroup-aware model development. Its primary significance lies in demonstrating that demographic bias in medical imaging can be mitigated without compromising diagnostic accuracy or image quality.

The authors emphasize that:

Fairness is achievable without a performance penalty: Subgroup equity can be improved while maintaining competitive predictive performance.
Clinical grounding is essential: Integrating physician-aligned preferences ensures that synthetic data preserves subtle, pathology-relevant features (e.g., cortical atrophy) that are often lost in standard generative models.
Scalability: The framework offers a scalable alternative to expensive retraining or fragile post-hoc controls.

The authors explicitly state that FairGen is not a deployment-stage decision-support system ready to replace existing workflows. Instead, it serves as a tool for dataset augmentation, educational illustration, and developing more equitable AI models. They acknowledge limitations, including the lack of formal multi-rater agreement statistics in the expert study and the need for prospective clinical validation, calibration, and risk analysis before clinical use. The work is presented as a step toward more equitable pre-deployment medical AI research.