ERDES: A Benchmark Video Dataset for Retinal Detachment and Macular Status Classification in Ocular Ultrasound.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Retinal detachment (RD) is a critical condition requiring urgent treatment.
• Macular status significantly impacts RD prognosis and surgical prioritization.
• Point-of-care ultrasound (POCUS) aids RD detection but requires expert interpretation.
• Current deep learning models lack clinical availability and do not assess macular status.

Purpose of the Study:

• To introduce the first open-access dataset (ERDES) for ocular ultrasound videos.
• To enable machine learning development for automated RD detection and macular status classification.
• To provide baseline performance benchmarks for deep learning models on ocular ultrasound data.

Main Methods:

• Developed the Eye Retinal DEtachment ultraSound (ERDES) dataset with labeled ocular ultrasound clips.
• Labeled data for the presence of RD and macular involvement (detached vs. intact).
• Trained and evaluated 40 baseline models across eight architectures, including 3D CNNs and transformers.

Main Results:

• The ERDES dataset is the first to support macular-based RD classification using ultrasound video.
• Baseline models were trained, demonstrating the feasibility of deep learning for RD detection and classification.
• Established benchmarks for future research in automated ocular ultrasound analysis.

Conclusions:

• The ERDES dataset facilitates the development of AI tools for automated RD diagnosis.
• Automated analysis of ocular ultrasound can improve diagnostic accuracy and efficiency.
• This work addresses a critical gap in AI-powered ophthalmology, particularly for resource-limited settings.