Artificial intelligence for epiretinal membrane detection using ultrawidefield scanning laser ophthalmoscopy

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• No validated AI model exists for epiretinal membrane (ERM) detection using ultrawidefield scanning laser ophthalmoscopy (UWF-SLO).
• This gap necessitates the development of automated tools for ERM screening.

Purpose of the Study:

• To develop and validate a deep learning (DL) model for automated ERM detection on UWF-SLO images.
• To assess the model's performance against human experts and evaluate AI-assisted diagnosis.

Main Methods:

• An optimized Inception-v3 DL model was trained on 920 UWF-SLO images, comparing macular and posterior pole regions.
• Models were validated internally and externally, benchmarked against ophthalmologists, and tested in an AI-assisted reader study.

Main Results:

• The macula AI model achieved an AUC of 0.943 (specificity 93.1%, sensitivity 84.0%), and the posterior pole model achieved an AUC of 0.931 (specificity 88.6%, sensitivity 86.0%).
• Both models demonstrated strong performance on external validation sets (AUCs 0.937 and 0.949).
• AI assistance improved ophthalmologists' diagnostic accuracy from 75.4% to 83.3%.

Conclusions:

• The study presents the first well-validated DL model for ERM detection on UWF-SLO images.
• The model exhibits robust performance in both internal and external validation.
• This AI tool can potentially enhance clinical screening and detection of ERMs.