Deep Learning Frameworks for Decision Support Systems in Ophthalmology

Abstract

Artificial intelligence (AI) is revolutionizing ophthalmology by enabling automated analysis of fundus and OCT images. This study evaluated convolutional neural networks (CNNs) for disease classification across multiple modalities. An ensembled CNN based on VGG16 achieved near-perfect sensitivity <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$(\approx 0.99-1.0)$</tex> and high specificity <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$(\approx 0.91-0.98$</tex> for cataract, glaucoma, and diabetic retinopathy detection from 4,217 fundus images, outperforming benchmark models. On an OCT dataset (>83,000 images), VGG16 and InceptionV3 achieved accuracy above 90%, with InceptionV3excelling in Normal and DME classification while VGG16 performed competitively on CNV and Drusen. Confusion matrices confirmed robust performance with minimal misclassification. Importantly, platforms such as IDHea, which provide access to curated, real-world ocular imaging datasets under strong governance, offer a path to addressing dataset diversity and enabling external validation. These findings demonstrate that CNN-driven AI systems are reliable, scalable tools for early detection and decision support, with strong potential for integration into clinical ophthalmology.

Ähnliche Arbeiten