Development and Evaluation of a Deep Learning Algorithm to Differentiate Between Membranes Attached to the Optic Disc...

Area of Science:

• Ophthalmology and Artificial Intelligence
• Medical Imaging Analysis
• Deep Learning in Healthcare

Background:

• Accurate differentiation of optic disc membranes, such as retinal detachment (RD) and posterior vitreous detachment (PVD), is crucial for timely patient management.
• Ocular ultrasonography (USG) is a valuable tool for visualizing these structures, but interpretation can be challenging in high-volume settings.
• The need for automated, reliable diagnostic aids is increasing in ophthalmology.

Purpose of the Study:

• To develop and validate a deep learning algorithm for identifying and differentiating between membranes attached to the optic disc (OD) using ocular USG.
• To specifically distinguish between posterior vitreous detachment (PVD) and retinal detachment (RD) based on B-scan ultrasonography images.

Main Methods:

• A transformer-based Vision Transformer (ViT) model, pre-trained on ImageNet21K, was utilized for image classification.
• Ocular B-scan ultrasonography images were pre-processed and classified into healthy, RD, or PVD categories.
• The dataset was split into training/validation (505 samples) and testing (212 samples) subsets to evaluate model performance.

Main Results:

• The AI model achieved high classification accuracy: 98.21% for PVD, 97.22% for RD, and 95.83% for normal cases.
• High sensitivity and specificity were reported for all categories, with specificity reaching 100% for RD.
• Minor misclassifications occurred, with seven instances of RD incorrectly identified as PVD.

Conclusions:

• A transformer-based deep learning algorithm demonstrates robust performance in distinguishing RD from PVD using ocular USG.
• The model enhances diagnostic efficiency, potentially leading to faster referrals and improved patient outcomes in urgent care.
• This AI innovation shows promise for clinical adoption in high-volume ocular imaging centers.