usama10/retinal-dr-longitudinal

Retinal DR Longitudinal: Paired Baseline and 2-Year Follow-up Fundus Images

Paired retinal fundus photographs from 574 diabetic retinopathy (DR) patients, with baseline and two-year follow-up images plus clinical metadata.

Dataset Description

This dataset contains longitudinal retinal fundus image pairs acquired at Tianjin Medical University Eye Hospital. Each patient has a baseline fundus photograph and a corresponding two-year follow-up photograph, enabling research on disease progression prediction, longitudinal image synthesis, and temporal retinal change analysis.

Ethics: This study adhered to the Declaration of Helsinki and was approved by the institutional review board of Tianjin Medical University Eye Hospital (approval number 2019KY-22). Written informed consent was obtained from all participants.

Dataset Statistics

Clinical Metadata

The file Organized_Data of Patients.xlsx contains two sheets (Baseline, 2-Year Follow-up) with the following variables per patient:

Grades 6 (post-photocoagulation) and 7 (ungradable) are present in the metadata but should be excluded from modeling due to non-standard retinal appearance or data incompleteness.

Directory Structure

baseline fundus images/
    00194/
        00194-7256.jpg      # Left or right eye
        00194-7261.jpg      # Other eye (if available)
    00197/
        ...
    (624 patient folders)

2 year follow-up fundus images/
    1036_2.jpg              # Flat directory, patient ID in filename
    1036_4.jpg
    ...
    (1,178 images)

Organized_Data of Patients.xlsx   # Clinical metadata
corrected_manifest.csv            # Corrected baseline<->follow-up eye pairing (use this)

Matching Baseline to Follow-up

Baseline images are organized in folders named by patient ID. Follow-up images are in a flat directory with filenames containing the patient ID as the prefix before the underscore. Some follow-up filenames have timestamp suffixes (_YYYYMMDD_HHMMSS) that should be stripped when parsing. Some baseline filenames have leading-zero inconsistencies; match by folder name rather than filename.

Corrected Eye Pairing (important)

Filenames encode the patient identifier but not a consistent eye-laterality marker, so matching baseline and follow-up images by filename order does not preserve which eye is which. For patients with two imaged eyes, this ordering is effectively independent of eye identity, and 258 of 572 patients (45.1%) end up with the baseline of one eye paired to the follow-up of the contralateral eye, an anatomically unrelated target that makes pixel-level prediction ill-posed.

We provide corrected_manifest.csv, which recovers the correct same-eye correspondence by image-based optimal assignment (maximizing mutual registration quality over the CLAHE-enhanced green channel). The recovery is validated independently at 95.5% agreement against an optic-disc laterality signal. We strongly recommend using this manifest rather than filename-order pairing. Its columns are:

Recommended Data Split

We recommend an 80/10/10 train/validation/test split at the patient level (both eyes of the same patient must be in the same split), stratified by DR progression label.

Associated Paper

Muhammad Usama, Emmanuel Eric Pazo, Xiaorong Li, Juping Liu. "Conditional Latent Diffusion for Predictive Retinal Fundus Image Synthesis from Baseline Imaging and Clinical Metadata." Manuscript, 2026 (under review).

Code: github.com/Usama1002/retinal-diffusion

Model weights: huggingface.co/usama10/retinal-diffusion-model

Citation

@article{usama2026retinal,
  title={Conditional Latent Diffusion for Predictive Retinal Fundus Image Synthesis from Baseline Imaging and Clinical Metadata},
  author={Usama, Muhammad and Pazo, Emmanuel Eric and Li, Xiaorong and Liu, Juping},
  note={Manuscript under review},
  year={2026}
}

License

This dataset is released under CC BY-NC 4.0. It may be used for non-commercial research purposes only. Redistribution of the raw images outside this repository is not permitted without explicit permission from the data providers.