Datasetpublic

doronys/synthmed-retina-hpp-sample

SynthMed Retina — HPP I2I Sample A sample of 2,500 synthetic retinal fundus images generated via image-to-image (I2I) diffusion from the Human Phenotype Project (10K) cohort, paired with anonymized patient metadata. Images were produced with the hpp-retina LoRA adapter at a fixed I2I strength of 0.5, which trades off structural fidelity against re-identification risk (see the paper for the full privacy–utility curve). Columns Column Type Description… See the full description on the dataset page: https://huggingface.co/datasets/doronys/synthmed-retina-hpp-sample.

sourceHugging Facecc-by-4.0updated 5mo agoView on Hugging Face

0likes35downloads

Dataset Card

SynthMed Retina — HPP I2I Sample

A sample of 2,500 synthetic retinal fundus images generated via image-to-image (I2I) diffusion from the Human Phenotype Project (10K) cohort, paired with anonymized patient metadata.

Images were produced with the `hpp-retina` LoRA adapter at a fixed I2I strength of 0.5, which trades off structural fidelity against re-identification risk (see the paper for the full privacy–utility curve).

Columns

Column	Type	Description
`image`	Image	512 × 512 RGB retinal fundus image
`name`	str	Unique image identifier
`strength`	float	I2I diffusion strength used for generation (always `0.5`)
`prompt`	str	Text prompt passed to the diffusion model
`participant_id`	int	Anonymized person ID — consistent across rows for the same participant
`eye`	str	Laterality — `"od"` (right) or `"os"` (left)
`age`	float	Age in years
`gender`	float	`0` = female, `1` = male
`bt__mcv`	float	Mean corpuscular volume (fL)
`bt__wbc`	float	White blood cell count (10³/µL)
`bt__platelets`	float	Platelet count (10³/µL)
`bt__hemoglobin`	float	Hemoglobin (g/dL)
`bt__creatinine`	float\	null	Creatinine (mg/dL)
`bt__glucose`	float\	null	Fasting glucose (mg/dL)
`bt__total_cholesterol`	float\	null	Total cholesterol (mg/dL)
`bt__hdl_cholesterol`	float\	null	HDL cholesterol (mg/dL)
`bt__non_hdl_cholesterol`	float\	null	Non-HDL cholesterol (mg/dL)
`bt__triglycerides`	float\	null	Triglycerides (mg/dL)
`bt__hba1c`	float\	null	HbA1c (%)
`bt__ferritin`	float\	null	Ferritin (ng/mL)
`bt__albumin`	float\	null	Albumin (g/dL)
`sitting_blood_pressure_pulse_rate`	float	Pulse rate (bpm)
`sitting_blood_pressure_systolic`	float	Systolic blood pressure (mmHg)
`bmi`	float	Body mass index (kg/m²)
`hips`	float	Hip circumference (cm)
`waist`	float	Waist circumference (cm)
`hand_grip_right`	float	Right-hand grip strength (kg)
`_norm` / `_cat`	float	Min-max normalized counterpart (`-1.0` = missing)

Note on missing values: raw label columns use null; the normalized *_norm columns use -1.0 as a sentinel for missing values.

Note on gender encoding: gender=0 → female, gender=1 → male (consistent with gender_cat class index 0 = female, 1 = male).

Generation Details

Parameter	Value
Base model	`CompVis/stable-diffusion-v1-4`
LoRA adapter	`doronys/synthmed-loras` — `hpp-retina/lora_weights.safetensors`
LoRA rank / alpha	64 / 32
Scheduler	DDIM
Inference steps	50
I2I strength	0.5
Resolution	512 × 512
Sample size	2,500 (one synthetic image per source image)

Source images are retinal fundus photographs from the HPP 10K cohort. Each source image was processed once (unique source → unique synthetic image).

Quick Start

python

from datasets import load_dataset

ds = load_dataset("doronys/synthmed-retina-hpp-sample", split="train")

# Each example has an 'image' (PIL.Image) plus all metadata columns
example = ds[0]
print(example["prompt"])   # "retina fundoscopy right eye dilated age=47 gender=female bp systolic=105"
print(example["age"])      # 47.0
example["image"].show()

About

These images were produced as part of a study evaluating I2I diffusion as a practical approach to medical image de-identification. Key findings:

—I2I outperforms T2I on pixel/perceptual fidelity and biomarker agreement at all conditioning strengths.
—At strength = 0.5: top-1 re-identification rate drops to ~20% while biomarker agreement (hemoglobin, Pearson r) remains ~0.7.
—Cross-cohort transfer: classifiers pretrained on I2I synthetic retinal images perform comparably to real-image pretraining when fine-tuning on UK Biobank data.

See the model card for the full privacy–utility curve and generation instructions.

Citation

bibtex

@article{synthmed2026,
  title   = {Privacy-Preserving Synthetic Medical Images via Image Conditioned Diffusion Models},
  author  = {Yaya-Stupp, Doron and Lutsker, Guy and Spiegel, Or and Segal, Eran},
  year    = {2026},
}

License

CC BY 4.0 — free to use, share, and adapt with attribution.

The underlying source images are from the Human Phenotype Project 10K cohort and are not included in this repository. Only the synthetic outputs and anonymized metadata are released here.