Development and validation of an offline deep learning algorithm to detect vitreoretinal abnormalities on ocular...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Ocular ultrasound (OUS) is crucial for diagnosing vitreoretinal abnormalities (VRA).
• Accurate and accessible diagnostic tools are needed, especially in underserved areas.

Purpose of the Study:

• To develop and validate an offline deep learning algorithm (DLA) for identifying VRA on OUS images.
• To assess the diagnostic performance of the DLA.

Main Methods:

• A dataset of 4740 OUS images was collected and annotated by vitreoretinal specialists.
• The DLA was trained on 4319 images and validated on 421 images.
• Performance metrics included sensitivity, specificity, PPV, NPV, and AUROC.

Main Results:

• The DLA achieved high sensitivity (90.8%) and specificity (97.1%) for VRA detection.
• Positive Predictive Value (PPV) was 97.0% and Negative Predictive Value (NPV) was 90.8%.
• The Area Under the Receiver Operating Characteristic Curve (AUROC) was 0.939, with excellent intergrader agreement (Cohen's kappa = 0.938).
• The DLA showed 100% sensitivity for vitreous hemorrhage and choroidal detachment, and 97.4% for retinal detachment.

Conclusions:

• The offline DLA demonstrates reliable performance for VRA detection on OUS.
• This algorithm can serve as a valuable tool for ophthalmic technicians in community eye screening, particularly in rural settings lacking specialists.