A Dataset of Widefield Fundus Images From Patients With Birdshot Chorioretinitis and Matched Control Authors: Thomas Foulonneau, Caroline Memmi, Dominique Monnet, Antoine P Brézin, Aliénor Vienne-Jumeau Published: 2026 ABSTRACT AbstractBirdshot chorioretinitis is a rare uveitis where research is limited by the lack of large, publicly available imaging datasets. To address this gap, we present Birdshot-Wide, a standardized retinal imaging dataset comprising 5,042 widefield fundus photographs from 742 eyes with birdshot chorioretinitis (BSCR) and 1,310 images from 742 age-and sex- matched healthy eyes controls. The BSCR cohort is longitudinal, with a median follow-up of 4.31 years. Each image in the BSCR cohort was manually reviewed for quality and classified by dominant lesion subtype. To demonstrate its utility, the dataset was benchmarked on a classification task, where deep learning models trained to differentiate BSCR from control images achieved an area under the receiver operating characteristic curve of up to 0.96. This dataset, with its detailed metadata, provides a valuable resource for biomarker discovery, the development of automated diagnostic tools, and disease progression modeling. Background & Summary Fundus photography is a cornerstone of ophthalmology for diagnosing and monitoring retinal diseases. For prevalent conditions such as diabetic retinopathy, numerous public image repositories have accelerated research and the development of automated analysis algorithms. However, for rare diseases, progress is often impeded by the scarcity of large, structured datasets1. Birdshot chorioretinitis (BSCR) is a rare, chronic form of posterior uveitis characterized by distinctive hypopigmented fundus lesions2. The automated analysis of these lesions could aid in diagnosis and monitoring, but the development of such tools requires substantial annotated data, which is currently lacking3. To address this gap, we present a large, curated dataset of widefield fundus images from patients with BSCR and matched healthy controls, collected at the uveitis clinic of Cochin Hospital (AP-HP, Paris). The patient data are derived from the prospective CO-BIRD cohort, which has previously been used to study the long-term evolution of visual function in patients with BSCR4. This dataset contains 5,042 images from 742 eyes with BSCR, making it the largest known curated imaging collection for this condition. Its longitudinal nature, with a median follow-up exceeding four years, allows for the study of disease progression over time. All images from the BSCR cohort were manually reviewed for quality and classified by dominant lesion subtype, providing a direct resource for supervised machine learning tasks. The dataset also includes 1,310 images from a cohort of 742 demographically matched healthy eyes. This control group provides a baseline for comparative analysis and is essential for the training and validation of supervised models for diagnostic classification. The dataset is available under controlled access on Zenodo (10.5281/zenodo.19474623)5. Prospective users must submit an access request and agree to the Data Usage Agreement, publicly visible on the repository page (see Usage Notes for details). Methods Data collection and ethics The dataset was constructed through a retrospective review of fundus photographs from the CO-BIRD cohort (ClinicalTrials.gov Identifier: NCT05153057), a prospective, single-center, observational cohort at Hôpital Cochin (AP-HP, Paris, France). This cohort, initiated in 2002, includes patients who are offered a standardized yearly evaluation. The study was approved by the Comité Consultatif de Protection des Personnes dans la Recherche Biomédicale (ethics approval no. 00011558) and conducted in accordance with the tenets of the Declaration of Helsinki. The ethics committee approved both the study protocol and the sharing of pseudonymized data. All participants provided written informed consent for participation in the CO-BIRD cohort and for the sharing of their pseudonymized imaging data for research purposes. All data were irreversibly pseudonymized prior to analysis, in full compliance with the General Data Protection Regulation (GDPR). Patient selection Patients from the CO-BIRD cohort were included in the Birdshot cohort if they had a confirmed diagnosis of birdshot chorioretinitis based on criteria defined by the International Consensus Conference and the Standardization of Uveitis Nomenclature (SUN) working group, including characteristic fundoscopic features and HLA-A29 positivity6. A mandatory inclusion criterion for the present dataset was the availability of at least one red–green image acquired by ultrawidefield retinal photography (Optos Silverstone, Optos PLC, Dunfermline, UK, colour red–green composite, ~200° field of view) in the hospital’s Picture Archiving and Communication System (PACS). The control cohort consisted of individuals imaged at the same institution who had no history of systemic autoimmune disease and whose included eye showed no evidence of retinal pathology, as confirmed manually and independently by two blinded retina specialists (AVJ, DM). Participants with unilateral retinal abnormalities were retained, but only the anatomically normal eye was included in the dataset. To ensure meaningful comparison, control eyes were matched to the Birdshot cohort based on sex and age distribution, stratified in ten-year age intervals. Image processing and standardization All images were retrieved from the institutional PACS in their native DICOM format. The workflow performed the extraction of relevant DICOM header metadata, including patient and study identifiers, demographic information, and image acquisition parameters. Concurrently, the raw pixel data were converted to and stored in the JPEG format, preserving the original base filename of the source DICOM file. This process results in non-identifying filenames. All extracted metadata, including the original filename, were compiled into a single structured XLSX file. This file serves as the essential lookup table to map each anonymized image file to its corresponding patient identifier, acquisition date, and laterality. Computational cohort matching A case-control matching procedure was implemented computationally to generate a demographically comparable control cohort. The metadata from the control subjects, initially organized at the image level, were first aggregated to create a patient-centric summary table. Each entry in this table represented a unique control subject, summarizing their demographic data and the availability of unilateral (OD or OS) or bilateral imaging. Subsequently, an iterative algorithm matched BSCR patients to control subjects from this aggregated pool. For each patient in the BSCR cohort, a corresponding control was sought based on two strict criteria: identity of sex and an age difference of no more than 10 years. The algorithm employed a hierarchical search strategy. Priority was given to identifying a single control subject with available imaging for both eyes. Should such a candidate not be available, a composite match was sought, consisting of two distinct control subjects—one with only right-eye imaging and another with only left-eye imaging—both independently satisfying the matching criteria. To ensure the statistical independence of the matched pairs, each control subject was selected at most once and was removed from the pool of available candidates upon being matched. Lesion classification and annotation An image was classified as ungradable if it fell into one of the following categories: (1) media opacities, including cataract and vitreous opacities, impairing visualization of more than 20% of the posterior pole; (2) poor centration, or eyelash/eyelid obscuration of the central retinal area extending to the superior or inferior vascular arcade; (3) presence of confounding retinal pathology (e.g., exudates, vasculitis, laser scars) that may obscure or mimic the lesions of interest. When multiple images were acquired for the same eye on the same examination date, all images were retained in the dataset; the metadata file records each image individually, allowing end-users to apply their own inclusion criteria. The remaining gradable images were assigned a dominant lesion type according to a predefined categorization scheme. Lesions were classified by retinal and uveitis specialists (AB, DM, AVJ) into morphological categories based on clinical presentation, including isolated or associated creamy spots and plaques, isolated plaques, pigmented spots and plaques, and atrophic spots. Additionally, specific phenotypes of pseudo-albinism and pseudo-melanocytosis were recorded7,8. To illustrate the potential utility of the dataset in segmentation and phenotyping tasks, a selected subset of images underwent manual annotation. Lesion contours were delineated using a vector-based tool9, and annotations were exported in JSON format along with summary statistics such as lesion count and area. These examples are provided as demonstration material for downstream computational applications. Data Record The Birdshot-Wide dataset is hosted on Zenodo as a controlled-access repository (10.5281/zenodo.19474623)5. As shown in Fig. 1, the unzipped file contains three subdirectories — BSCR/, Controls/, and Annotations/ — alongside a single metadata file, metadata.xlsx. The Annotations/ subdirectory contains the lesion annotation files in JSON format. Each annotation file is named to match its corresponding source image (e.g., the annotation for “001.jpg” is stored as “001.json”). Data records root structure of Birdshot-WIDE. The BSCR/ and Controls/ directories contain a total of 6,352 ultra-widefield fundus images in JPEG format. The filenames of the images are non-descriptive and correspond to their original source DICOM filenames. Consequently, all linkage between an image file and its corresponding clinical and demographic data must be performed using the metadata.xlsx file. This metadata file provides a comprehensive record for each image in the dataset. The data fields are defined as follows: filename: The unique, non-identifying filename of the image, corresponding to the files in the BSCR/ and Controls/ directories. filename: The unique, non-identifying filename of the image, corresponding to the files in the BSCR/ and Controls/ directories. patient_id: A pseudonymized patient identifier, which allows for the grouping of all images belonging to a single individual. patient_id: A pseudonymized patient identifier, which allows for the grouping of all images belonging to a single individual. acquisition_date: The date of the exam on which the image was acquired. acquisition_date: The date of the exam on which the image was acquired. age_at_acquisition_date: The patient’s age at the time the image was acquired. age_at_acquisition_date: The patient’s age at the time the image was acquired. sex: The patient’s sex (‘Male’ or ‘Female’). sex: The patient’s sex (‘Male’ or ‘Female’). laterality: The laterality of the imaged eye (‘R’ for Right or ‘L’ for Left) laterality: The laterality of the imaged eye (‘R’ for Right or ‘L’ for Left) lesion_type: A categorical label describing the dominant lesion pattern as determined by manual review. Although multiple lesion types may co-exist within a single image, each image was assigned a single label reflecting the predominant morphological pattern. This field is populated for the BSCR cohort. Possible labels include ‘Excluded’ (for images deemed of insufficient quality), ‘Pseudo-albinism’, ‘Pseudo-melanocytosis’, ‘Pigmented plaque and spots’, ‘Pigmented spots without plaque’, ‘Non pigmented plaque and spots’, ‘Non pigmented plaque’, ‘Non pigmented atrophic and non-atrophic spots’, and ‘Non pigmented non atrophic spots’. lesion_type: A categorical label describing the dominant lesion pattern as determined by manual review. Although multiple lesion types may co-exist within a single image, each image was assigned a single label reflecting the predominant morphological pattern. This field is populated for the BSCR cohort. Possible labels include ‘Excluded’ (for images deemed of insufficient quality), ‘Pseudo-albinism’, ‘Pseudo-melanocytosis’, ‘Pigmented plaque and spots’, ‘Pigmented spots without plaque’, ‘Non pigmented plaque and spots’, ‘Non pigmented plaque’, ‘Non pigmented atrophic and non-atrophic spots’, and ‘Non pigmented non atrophic spots’. cohort: A binary categorical variable indicating the subject’s group (‘BSCR’ or ‘Control’). cohort: A binary categorical variable indicating the subject’s group (‘BSCR’ or ‘Control’). Technical Validation This section describes the characteristics of the dataset and the quality control steps taken to ensure its utility for research. Cohort composition and demographics The final Birdshot-Wide dataset comprises 5,042 color fundus photographs from 742 eyes with BSCR, each with at least one usable color fundus image after quality control5. The mean age of patients with BSCR was 66.16 ± 12.27 years, and 59.8% were female. On average, each eye contributed 6.07 ± 5.17 images, reflecting the longitudinal nature of care. The matched control cohort includes 742 eyes without known retinal disease, from which 1,310 color fundus images were retained (Fig. 2). Controls were age- and sex-matched to the BSCR cohort (mean age 63.78 ± 14.14 years; 59.8% female) (Table 1). Cohort composition and image selection workflow. More than 12,000 DICOM files were retrieved. After conversion, de-duplication, and quality control, 5,042 color fundus images from 742 BSCR eyes were retained. The matched control cohort includes 742 eyes contributing 1,310 color fundus images. Non-color modalities (e.g., fundus autofluorescence) were excluded. Table 1Demographic and dataset characteristics of the Birdshot and control cohorts.CharacteristicBSCR eyes (n = 742)Controls eyes (n = 742)Age, mean (SD)66.16 (12.27)63.78 (14.14)Male, n (%)298 (40.20)298 (40.20)Number of OD images (%)371 (50)371 (50)Images per eye, mean (SD)6.07 (5.17)1.46 (2.03)Follow-up (years), median [IQR]4.31 [0 .89–6.09]0.00 (0.00–0.00)BSCR: Birdshot chorioretinis Demographic and dataset characteristics of the Birdshot and control cohorts. BSCR: Birdshot chorioretinis Image quality and lesion classification All color fundus images in the BSCR cohort underwent manual review (n = 5,042). Of these, 816 images (16.2%) were deemed of insufficient quality—due to poor illumination, motion artifacts, or media opacities—and were flagged accordingly. Among the remaining gradable images, the most prevalent phenotypes were creamy plaques and spots (n = 2,262; 44.9%) and isolated creamy spots (n = 1,122; 22.3%). Other observed patterns included isolated plaques (n = 299; 5.9%), pigmented plaques and spots (n = 259; 5.1%), pigmented spots (n = 80; 1.6%), and atrophic spots (n = 43; 0.8%). Additionally, distinct appearances of pseudo-albinism (n = 131; 2.6%) and pseudo-melanocytosis (n = 30; 0.6%) were recorded. Each gradable image received a categorical lesion-type label, provided in the dataset metadata. Longitudinal data characteristics The dataset is longitudinal. In the BSCR cohort, the median eye-level follow-up was 4.31 years (IQR 0.89–6.09). A subset of eyes has serial fundus photographs spanning up to 7.32 years, enabling monitoring of lesion evolution over time. Figure 3 provides a representative example, illustrating progressive funduscopic changes in an eye over a seven-year period. Longitudinal follow-up of a patient with Birdshot Chorioretinitis over 7 years. Ultra-widefield color fundus photographs of the right (OD) and left (OS) eyes at baseline (T0) and after 7 years (T7). Progressive atrophic changes and redistribution of hypopigmented lesions are visible over time. Illustrative annotations for quantitative analysis To demonstrate the potential for quantitative image analysis, a subset of BSCR images underwent semi-automatic annotation. Five patients with longitudinal follow-up were selected, and atrophic lesions were delineated on ultra-widefield images at two time points (baseline and five-year follow-up). The mean number of annotated zones was 29.0 ± 22.9 at baseline and 27.8 ± 22.1 at the five-year follow-up. The mean total lesion area, calculated by summing the three lesion subtypes, was 68.1 kpixels² at baseline and 78.5 kpixels² after five years. Regarding the mean angular position relative to the optic disc, the value for typical creamy lesions was 9.0° ± 5.6° at baseline and -35.6° at follow-up. For atrophic lesions, the mean angle was -10.1° ± 12.6° at baseline and -3.8° ± 7.8° at the five-year mark (Table 2). These annotations are provided alongside the image dataset as example materials for use in lesion segmentation and disease quantification tasks. Table 2Quantitative lesion metrics at baseline and at 5-year follow-up.VariableBaseline (n = 5)5-year follow-up (n = 5)Number of annotated zones29.0 ± 22.927.8 ± 22.1Typical creamy BCR lesions Number of zones8.2 ± 17.20.2 ± 0.4 Total area (kpixels²)10.0 ± 21.80.5 ± 1.0 Mean area per zone (kpixels²)0.8 ± 0.52.3 Mean major axis (pixels)62.1 ± 16.4102.8 Mean minor axis (pixels)42.9 ± 4.867.0 Mean elongation index1.5 ± 0.31.5 Mean circularity index0.8 ± 0.10.8 Mean angle from ONH (°)9.0 ± 5.6-35.6 Mean ONH distance (pixels)365.6 ± 76.7582.4Atrophic BCR lesions Number of zones20.2 ± 25.126.8 ± 21.6 Total area (kpixels²)33.4 ± 37.241.3 ± 38.1 Mean area per zone (kpixels²)2.4 ± 0.91.7 ± 0.7 Mean major axis (pixels)108.4 ± 13.097.9 ± 24.3 Mean minor axis (pixels)60.8 ± 13.850.6 ± 9.3 Mean elongation index1.8 ± 0.31.9 ± 0.4 Mean circularity index0.8 ± 0.10.7 ± 0.1 Mean angle from ONH (°)−10.1 ± 12.6-3.8 ± 7.8 Mean ONH distance (pixels)504.0 ± 39.6513.5 ± 46.8Large atrophic plaques Number of zones0.6 ± 0.50.8 ± 0.8 Total area (kpixels²)24.7 ± 26.436.7 ± 33.5 Mean area per zone (kpixels²)41.1 ± 19.551.2 ± 18.6Abbreviations: BCR: Birdshot chorioretinopathy; ONH: optic nerve head. Quantitative lesion metrics at baseline and at 5-year follow-up. Abbreviations: BCR: Birdshot chorioretinopathy; ONH: optic nerve head. Available for AI training To demonstrate the technical efficacy of the Birdshot-Wide dataset and to provide a baseline for future work, we conducted a benchmark experiment on a supervised binary classification task. The objective was to train machine learning models to automatically distinguish fundus images of patients with BSCR from those of healthy controls. For this task, we employed four standard convolutional neural network (CNN) architectures, pre-trained on ImageNet: ResNet5010, VGG1611, EfficientNet-B012, and Densenet12113. For each model, the final classification layer was replaced with a fully connected layer adapted to our binary classification problem. To ensure a subject-independent evaluation, the dataset was partitioned at the patient level into a test set (20%) and a main training set (80%). This main training set was then further split to create a validation set (approx. 10% of the total data), which was used for hyperparameter tuning and early stopping. This partitioning guaranteed that no images from the same individual were present across the training, validation, or test subsets. Images were resized to 224 × 224 pixels. Standard data augmentation techniques were applied to the training set, including random resized cropping, horizontal flipping, color jitter, and minor random rotations. The network weights were updated up to 15 epochs by minimizing a CrossEntropyLoss function14. We employed the Adam optimization algorithm15 with a learning rate of 0.001 and a batch size of 64. To optimize performance, we utilized mixed-precision training16. Performance was assessed using the Area Under the Receiver Operating Characteristic Curve (AUC-ROC), accuracy, precision, recall, and F1-score. The quantitative results from the evaluation on the test set are presented in Table 3. The models demonstrated robust performance, with the RESNET50 architecture achieving the highest AUC-ROC of 0.96 and an overall accuracy of 89%. These results confirm the high utility of the Birdshot-Wide dataset for training differential diagnostic models. A qualitative analysis of classification errors indicated that the most challenging cases primarily involved low-quality images and early disease stages with subtle lesions, offering avenues for improvement in future algorithms (Fig. 4). Table 3Performance of Convolutional Neural Network models for Birdshot chorioretinopathy (BCR) classification on the test set.ModelAccuracyAUC-ROCPrecisionRecallF1-ScoreRESNET500.890.960.960.890.92VGG160.730.870.940.670.78EFFICIENTNETB00.830.910.940.820.88DENSENET1210.890.950.900.950.93BCR: Birdshot chorioretinopathy; AUC-ROC: Area Under the Receiver Operating Characteristic curve. Performance of Convolutional Neural Network models for Birdshot chorioretinopathy (BCR) classification on the test set. BCR: Birdshot chorioretinopathy; AUC-ROC: Area Under the Receiver Operating Characteristic curve. Examples of classification errors from the benchmark experiment. (A) A true BSCR image misclassified as control, associated with poor image quality and low contrast. (B) A control image misclassified as BSCR, showing significant image degradation due to cataract. (C) A true BSCR image misclassified as control, presenting with subtle lesions. (D) A control image misclassified as BSCR, obscured by a prominent eyelid artifact. Beyond classification, the lesion-level annotations provided for a subset of images pave the way for segmentation applications. This data can serve as a foundation for training and validating models capable of automatically delineating and quantifying lesions, thereby enabling the objective tracking of disease progression over time. Usage Notes The Birdshot-Wide dataset is shared under controlled access due to the sensitive nature of widefield fundus images, which carry a potential risk of participant re-identification. The dataset is hosted on Zenodo (10.5281/zenodo.19474623). Prospective users must submit an access request through the Zenodo platform and agree to the Data Usage Agreement (DUA), which is publicly visible on the repository page. Access requests are typically processed within five business days. Users are asked to indicate the general purpose of their request in their application; however, access is granted to all users who agree to the terms of the DUA, regardless of the research topic. The longitudinal structure of this dataset is particularly suited for developing and validating algorithms for disease progression modeling. Specifically, the presence of a large, matched control group makes this dataset ideal for training and testing supervised classification algorithms or for anomaly detection models. Researchers can use the serial images to track the evolution of lesions over time. The inclusion of images with suboptimal quality, which are clearly flagged in the metadata, allows for the creation of robustness-aware machine learning models that better reflect real-world clinical conditions. Furthermore, the provided illustrative annotations can serve as a starting point or as a validation set for developing and testing automated lesion segmentation and tracking applications. Limitations Several limitations should be acknowledged. First, the dataset was constructed from a single tertiary center, which may introduce selection bias in terms of disease severity or referral patterns. Second, despite efforts to standardize processing, image acquisition protocols varied across devices and time, potentially introducing technical heterogeneity. Third, lesion classification was based on manual visual assessment by trained graders, without ground-truth confirmation via multimodal imaging such as fluorescein angiography or OCT. Additionally, each image was assigned a single dominant lesion-type label, although multiple lesion types may co-exist within the same image; a multi-label annotation scheme could provide richer phenotypic information in future extensions of the dataset. Finally, although a control cohort was matched by age and sex, additional ophthalmic or systemic comorbidities could not be systematically excluded due to the retrospective nature of data extraction. Author contributions T.F. and A.V.J. conceptualized the paper, conceived the experiments and wrote the manuscript, T.F. and A.V.J. and C.M. did data curation. D.M. A.P.B. provided the original data. All authors reviewed the manuscript. Funding No specific funding: This research received no specific grant from funding agencies in the public, commercial, or not-for-profit sectors. Data availability The Birdshot-Wide dataset is available under controlled access on Zenodo (10.5281/zenodo.19474623). Prospective users must submit an access request through the platform and agree to the Data Usage Agreement, publicly visible on the repository page, to gain access. Code availability The custom code used for data processing, cohort matching, and the benchmark classification experiment is publicly available at https://github.com/thomasb6/Birdshot-WIDE. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement The Birdshot-Wide dataset is available under controlled access on Zenodo (10.5281/zenodo.19474623). Prospective users must submit an access request through the platform and agree to the Data Usage Agreement, publicly visible on the repository page, to gain access. The custom code used for data processing, cohort matching, and the benchmark classification experiment is publicly available at https://github.com/thomasb6/Birdshot-WIDE.