Retinal Image Database for ROP Plus Disease AI Diagnosis

A 2026 study presents a cross-center retinal image database containing 2,635 images for training AI models to diagnose retinopathy of prematurity (ROP) plus disease. The research compares nine deep learning architectures, including ResNet50 and Swin-Transformer, and reports model performance metrics like AUC, accuracy, and F1-score. The dataset was compiled by Xiqianru Zhang from public and private sources to test a multi-center fusion strategy.

Use Cases

• Train and compare the performance of nine deep learning models (e.g., ResNet50, Swin-Transformer) on 2,635 retinal images for ROP classification.
• Validate AI diagnostic models using reported metrics like AUC, accuracy, recall, and F1-score from cross-center testing.
• Implement the described cross-center fusion strategy to improve model generalizability for ROP plus disease detection.

Strengths

• Contains 2,635 retinal images compiled from multiple clinical centers, providing a foundation for testing model generalizability.
• Benchmarks nine distinct deep learning architectures with specific performance metrics (e.g., ResNet50 achieved an AUC of 1.00 for Plus disease).
• The cross-center fusion training strategy yielded a model with 92.60% accuracy, surpassing single-center results.

Limitations

• The dataset is described within a 128.1 KB PDF research paper; the underlying image data files are not directly accessible.
• Sample size is limited to 2,635 images, which may be insufficient for training robust models without extensive data augmentation.
• Specific image metadata, patient demographics, and detailed columnar data are unavailable from the provided description.

Provenance

Source

figshare, authored by Xiqianru Zhang.

Collection Method

Merged public datasets (FARFUM-RoP, HVDROPDB, LAN-RoP) and private sets from three tertiary NICUs.

Freshness

Last updated on 2026-03-24.