Retinal Images for Prematurity Plus Disease Diagnosis Across Multiple Centers

A research paper describes a cross-center retinal image database containing 2,635 images for diagnosing retinopathy of prematurity (ROP) plus disease. The study, by Xiqianru Zhang, compares nine deep learning models, with ResNet50 achieving the best performance metrics, including an AUC of up to 1.00 for Plus disease classification. The consolidated multi-center cohort yielded optimal model performance with an accuracy of 92.60%.

Use Cases

• Compare the performance of nine deep learning models (ResNet34, ResNet50, DenseNet, Inception, MobileNet, VGG16, VGG19, EfficientNet, Swin-Transformer) on retinal images for ROP diagnosis.
• Train a ResNet50 model on the 2,635-image cross-center database to classify Normal, Pre-plus, and Plus disease stages, targeting reported AUC values of 0.97, 0.95, and 1.00.
• Validate model generalizability by testing on a multi-center dataset, using metrics like accuracy, recall, precision, and F1-score, which reached 92.60% in the study.
• Analyze the impact of cross-center fusion versus single-center training on diagnostic indices such as FLOPs and parameter count for ROP screening models.

Strengths

• Contains 2,635 retinal images from a merged multi-center database, providing a basis for testing model generalizability.
• Paper reports high model performance, with ResNet50 achieving an AUC of 1.00 for Plus disease classification on the test set.
• Comparative analysis includes nine distinct deep learning architectures and multiple evaluation metrics (FLOPs, parameter count, accuracy, recall, precision, F1 score).

Limitations

• The dataset is described within a research paper (PDF); the underlying image files are not directly provided, limiting immediate use for model training.
• Sample size is limited to 2,635 images, which is small for training complex deep learning models from scratch without augmentation.
• Specific image metadata, patient demographics, and detailed labeling procedures are not described in the provided input.

Provenance

Source

Xiqianru Zhang via figshare.

Collection Method

Merged public and private retinal image sets (FARFUM-RoP, HVDROPDB, LAN-RoP, Preterm infants <34 weeks GA from three tertiary NICUs).

Freshness

Last updated on 2026-03-24.