Datasets

FairMedFM evaluates fairness on medical imaging datasets that record patient attributes. Several datasets require an application to their provider, so FairMedFM cannot redistribute them; the repository provides the preprocessing scripts instead.

Classification

--dataset	Dataset	Modality	Sensitive attributes	Access
CXP	CheXpert (demographics)	Chest X-ray	Sex, Age, Race	Application
MIMIC_CXR	MIMIC-CXR	Chest X-ray	Sex, Age, Race	Credentialed (PhysioNet)
HAM10000	HAM10000	Dermoscopy	Sex, Age	Preprocessed download
PAPILA	PAPILA	Fundus	Sex, Age	Preprocessed download
FairVLMed10k	FairVLMed10k	Fundus (SLO)	Sex, Age, Race, Language	Provider
GF3300	Harvard-GF (GF3300)	Fundus	Sex, Age, Race, Language	Provider
BREST	BRSET	Fundus	Sex, Age	Provider
COVID_CT_MD	COVID-CT-MD	CT	Sex, Age	Public
ADNI	ADNI 1.5T	Brain MRI	Sex	Application

Use --sensitive_name with one of the dataset's attributes (Sex, Age, Race or Language).

Segmentation

--dataset	Dataset	Modality
TUSC	TUSC	Thyroid ultrasound
HAM10000-Seg	HAM10000	Dermoscopy
FairSeg	Harvard-FairSeg	Fundus (SLO)
montgomery	Montgomery County X-ray	Chest X-ray

TUSC has a packaged config and records sex; the other segmentation datasets need a config in ./configs/datasets/ (see Run experiments). The paper also covers 3D segmentation datasets (KiTS2023, IRCADb, CANDI and SPIDER), whose preprocessing scripts are in the repository.

Preparing data

1. Download the data from the provider (or the preprocessed archive, where available) into data/<dataset>/ in your working directory.
2. Run the dataset's preprocessing notebook from pre-processing/. The notebooks preprocess images where needed, extract the sensitive attributes, and for classification create train/test splits with balanced subgroups.
3. Check that the paths in the dataset config match your files, or override the config in ./configs/datasets/.

Using your own dataset

To measure fairness on your own data with your own model, you do not need a dataset config: export predictions and use fairmedfm score.