FairGenMed is the first dataset for studying fairness in medical generative models. It provides detailed quantitative clinical measurements alongside demographic annotations to investigate the semantic correlation between text prompts and anatomical regions across demographic subgroups. The dataset supports both generative model evaluation and downstream classification tasks for glaucoma detection.
This dataset accompanies the FairDiffusion framework — an equity-aware latent diffusion model that enhances fairness in medical image generation via Fair Bayesian Perturbation — published in Science Advances (2025).
| Field | Value |
|---|---|
| Institution | Department of Ophthalmology, Harvard Medical School |
| Task | Glaucoma detection; fairness evaluation of generative models |
| Modality | Scanning Laser Ophthalmoscopy (SLO) fundus images, OCT B-scans |
| Scale | 10,000 subjects |
| Image size | 512 × 664 (SLO fundus) |
| License | CC BY-NC-ND 4.0 |
Each subject includes one SLO fundus image and one .npz file. The NPZ files contain:
| Field | Description |
|---|---|
glaucoma | Disease label: 0 = non-glaucoma, 1 = glaucoma |
oct_bscans | OCT B-scan images |
race | 0 = Asian, 1 = Black, 2 = White |
male | 0 = Female, 1 = Male |
hispanic | 0 = Non-Hispanic, 1 = Hispanic |
maritalstatus | 0 = Married/Partnered, 1 = Single, 2 = Divorced, 3 = Widowed, 4 = Legally Separated, -1 = Unknown |
language | 0 = English, 1 = Spanish, 2 = Other |
All clinical measurements for the 10,000 samples are provided in data_summary.csv:
| Column | Description |
|---|---|
cdr_status | Cup-disc ratio status |
md_severity | Severity of vision loss |
se_status | Spherical equivalent status |
6 demographic attributes are annotated per subject: age, gender, race, ethnicity, preferred language, and marital status.
Clinical decisions, patient care, or any commercial application. This dataset shall not be used for clinical decisions at any time.
FairDiffusion is an equity-aware latent diffusion model built on Stable Diffusion 2.1, trained with Fair Bayesian Perturbation to reduce demographic bias in generated medical images. It is evaluated on FairGenMed (ophthalmology), HAM10000 (dermatology), and CheXpert (chest radiology).
BibTeX:
@article{FairDiffusion_Science_Advances_2025,
author = {Yan Luo and Muhammad Osama Khan and Congcong Wen and Muhammad Muneeb Afzal and Titus Fidelis Wuermeling and Min Shi and Yu Tian and Yi Fang and Mengyu Wang},
title = {FairDiffusion: Enhancing equity in latent diffusion models via fair Bayesian perturbation},
journal = {Science Advances},
volume = {11},
number = {14},
pages = {eads4593},
year = {2025},
doi = {10.1126/sciadv.ads4593}
}
APA:
Luo, Y., Khan, M. O., Wen, C., Afzal, M. M., Wuermeling, T. F., Shi, M., Tian, Y., Fang, Y., & Wang, M. (2025). FairDiffusion: Enhancing equity in latent diffusion models via fair Bayesian perturbation. Science Advances, 11(14), eads4593. https://doi.org/10.1126/sciadv.ads4593
FairGenMed is the first dataset for studying fairness in medical generative models. It provides detailed quantitative clinical measurements alongside demographic annotations to investigate the semantic correlation between text prompts and anatomical regions across demographic subgroups. The dataset supports both generative model evaluation and downstream classification tasks for glaucoma detection.
This dataset accompanies the FairDiffusion framework — an equity-aware latent diffusion model that enhances fairness in medical image generation via Fair Bayesian Perturbation — published in Science Advances (2025).
| Field | Value |
|---|---|
| Institution | Department of Ophthalmology, Harvard Medical School |
| Task | Glaucoma detection; fairness evaluation of generative models |
| Modality | Scanning Laser Ophthalmoscopy (SLO) fundus images, OCT B-scans |
| Scale | 10,000 subjects |
| Image size | 512 × 664 (SLO fundus) |
| License | CC BY-NC-ND 4.0 |
Each subject includes one SLO fundus image and one .npz file. The NPZ files contain:
| Field | Description |
|---|---|
glaucoma | Disease label: 0 = non-glaucoma, 1 = glaucoma |
oct_bscans | OCT B-scan images |
race | 0 = Asian, 1 = Black, 2 = White |
male | 0 = Female, 1 = Male |
hispanic | 0 = Non-Hispanic, 1 = Hispanic |
maritalstatus | 0 = Married/Partnered, 1 = Single, 2 = Divorced, 3 = Widowed, 4 = Legally Separated, -1 = Unknown |
language | 0 = English, 1 = Spanish, 2 = Other |
All clinical measurements for the 10,000 samples are provided in data_summary.csv:
| Column | Description |
|---|---|
cdr_status | Cup-disc ratio status |
md_severity | Severity of vision loss |
se_status | Spherical equivalent status |
6 demographic attributes are annotated per subject: age, gender, race, ethnicity, preferred language, and marital status.
Clinical decisions, patient care, or any commercial application. This dataset shall not be used for clinical decisions at any time.
FairDiffusion is an equity-aware latent diffusion model built on Stable Diffusion 2.1, trained with Fair Bayesian Perturbation to reduce demographic bias in generated medical images. It is evaluated on FairGenMed (ophthalmology), HAM10000 (dermatology), and CheXpert (chest radiology).
BibTeX:
@article{FairDiffusion_Science_Advances_2025,
author = {Yan Luo and Muhammad Osama Khan and Congcong Wen and Muhammad Muneeb Afzal and Titus Fidelis Wuermeling and Min Shi and Yu Tian and Yi Fang and Mengyu Wang},
title = {FairDiffusion: Enhancing equity in latent diffusion models via fair Bayesian perturbation},
journal = {Science Advances},
volume = {11},
number = {14},
pages = {eads4593},
year = {2025},
doi = {10.1126/sciadv.ads4593}
}
APA:
Luo, Y., Khan, M. O., Wen, C., Afzal, M. M., Wuermeling, T. F., Shi, M., Tian, Y., Fang, Y., & Wang, M. (2025). FairDiffusion: Enhancing equity in latent diffusion models via fair Bayesian perturbation. Science Advances, 11(14), eads4593. https://doi.org/10.1126/sciadv.ads4593