AI-Assisted Interpretation of Ultra-Widefield Retinal Images

This study is a prospective multi-center observational reader study designed to evaluate the impact of artificial intelligence (AI) assistance on clinician interpretation of ultra-widefield (UWF) retinal images.

Approximately 600 UWF retinal images will be prospectively collected from multiple ophthalmic centers in China. Images will be acquired using clinically routine UWF retinal imaging systems and will include a broad spectrum of retinal diseases and retinal findings encountered in real-world clinical practice.

All images will undergo independent expert annotation by retinal specialists to establish reference labels for retinal finding categories. These expert annotations will serve as the reference standard for subsequent performance evaluation.

Four ophthalmologists with different levels of clinical experience will participate in the reader study, including:

one senior retinal specialist with approximately five years of retinal clinical experience; three junior ophthalmologists with approximately two years of ophthalmology residency training.

A randomized crossover multi-reader design will be implemented to minimize recall bias and balance reading conditions.

For each clinician, the image dataset will be randomly divided into two equal subsets (subset A and subset B; approximately 300 images each).

During Round 1:

subset A will be interpreted without AI assistance; subset B will be interpreted with AI assistance.

After a washout interval of at least two weeks, the reading conditions will be reversed during Round 2:

subset A will be interpreted with AI assistance; subset B will be interpreted without AI assistance.

Image order will be independently randomized for each session and each clinician.

Under the unaided condition, clinicians will evaluate retinal images using standard clinical interpretation without AI output.

Under the AI-assisted condition, clinicians will receive category-level AI prediction probabilities for retinal finding categories. The AI output will provide probabilistic confidence scores only and will not include lesion localization maps, heatmaps, segmentation overlays, or automated binary recommendations.

Clinicians will remain blinded to the expert reference labels and to the interpretations of other readers. Final diagnostic decisions will be independently determined by each clinician.

The primary analysis will compare diagnostic performance between unaided and AI-assisted conditions, including sensitivity, specificity, area under the receiver operating characteristic curve (AUC), and inter-reader agreement.