- Python 100%
| Filename | Latest commit message | Latest commit date |
|---|---|---|
| benchmark | ||
| data_processing | ||
| figures | ||
| matching | ||
| LICENSE | ||
| README.md | ||
| requirements.txt | ||
Birdshot-WIDE
A Dataset of Widefield Fundus Images From Patients With Birdshot Chorioretinitis and Matched Control
Overview
This repository contains the custom code used for data processing, cohort matching, and the benchmark classification experiment described in the associated data descriptor published in Scientific Data.
The Birdshot-WIDE dataset comprises 5,042 widefield fundus photographs from 742 eyes with birdshot chorioretinitis (BSCR) and 1,310 images from 742 age- and sex-matched healthy control eyes, acquired using the Optos Silverstone ultra-widefield scanning laser ophthalmoscope.
Repository Structure
birdshot-wide/
├── data_processing/
│ ├── 01_dicom_extraction.py # DICOM to JPEG conversion + metadata extraction
│ ├── 02_patient_aggregation.py # Aggregate image-level metadata to patient level
│ ├── 03_add_cohort_label.py # Add cohort column (BCR / Control)
│ ├── 04_add_lesion_type.py # Assign lesion-type labels from sorted folders
│ └── 05_anonymize_metadata.py # Pseudonymize patient identifiers (SHA-256)
├── matching/
│ └── case_control_matching.py # Iterative age- and sex-matched case-control pairing
├── benchmark/
│ └── classification.py # CNN benchmark (ResNet50, VGG16, EfficientNet-B0, DenseNet121)
├── figures/
│ ├── table1_demographics.py # Generate Table 1 (cohort demographics)
│ └── figure3_lesion_distribution.py # Lesion-type frequency plot
├── requirements.txt
├── LICENSE
└── README.md
Pipeline
The scripts are numbered in the order they should be executed:
| Step | Script | Description |
|---|---|---|
| 1 | data_processing/01_dicom_extraction.py |
Extract DICOM files from PACS, convert to JPEG, and generate image-level metadata |
| 2 | data_processing/02_patient_aggregation.py |
Aggregate metadata to one row per patient with image counts per eye |
| 3 | matching/case_control_matching.py |
Match BSCR patients to controls by sex and age (±10 years) |
| 4 | data_processing/03_add_cohort_label.py |
Add the cohort column to control metadata |
| 5 | data_processing/04_add_lesion_type.py |
Assign dominant lesion-type labels based on manual sorting |
| 6 | data_processing/05_anonymize_metadata.py |
Generate pseudonymized patient IDs and produce the final metadata.xlsx |
| 7 | benchmark/classification.py |
Train and evaluate CNN models for BSCR vs. Control classification |
Requirements
pip install -r requirements.txt
Key dependencies:
- Python ≥ 3.9
- PyTorch ≥ 2.0
- torchvision
- pydicom
- pandas, openpyxl
- scikit-learn
- Pillow
- matplotlib, seaborn
- dcmtk (system package for
dcmj2pnm)
Dataset Access
The dataset is available under controlled access on Zenodo:
https://doi.org/10.5281/zenodo.19474623
Prospective users must submit an access request through the Zenodo platform and agree to the Data Usage Agreement (DUA), publicly visible on the repository page.
Annotation Tool
Lesion annotations were created using FundusTracker, a vector-based annotation tool for widefield fundus images.
Citation
If you use this dataset or code, please cite:
@article{foulonneau2026birdshot,
title={A Dataset of Widefield Fundus Images From Patients With Birdshot Chorioretinitis and Matched Control},
author={Foulonneau, Thomas and Memmi, Caroline and Monnet, Dominique and Br{\'e}zin, Antoine P. and Vienne-Jumeau, Ali{\'e}nor},
journal={Scientific Data},
year={2026},
doi={10.5281/zenodo.19474623}
}
License
This code is released under the MIT License.