Deep Learning-Based Classification of OCT Images for Retinal Disease Detection

Resumo

Retinal diseases, such as Age-Related Macular Degeneration (AMD), Diabetic Macular Edema (DME), and Choroidal Neovascularization (CNV), are leading causes of vision impairment globally. These conditions are characterized by structural alterations in the retina, including the formation of deposits like drusen, fluid accumulation, and abnormal blood vessel growth, which can progressively impair vision if not detected early. Optical Coherence Tomography (OCT) is a non-invasive imaging technique that provides high-resolution cross-sectional images of the retina, enabling detailed visualization of these pathological features.This study focuses on developing and evaluating deep neural network (DNN) models for the automatic classification of OCT images aimed at detecting these ocular diseases. Using transfer learning, the research adapts pretrained architectures, including VGG16, VGG19, InceptionV3, ResNet50, and DenseNet121, to specialize in identifying disease-specific features within OCT scans. The models are trained on a balanced dataset, utilizing data augmentation and dataset balancing strategies to enhance their generalization and robustness.Performance metrics such as accuracy, sensitivity, specificity, and Area Under the Receiver Operating Characteristic Curve (AUC) demonstrate that the models achieve high classification effectiveness. Further analysis through confusion matrices and activation maps indicates that the networks focus on clinically relevant regions within the images—such as areas of drusen, fluid accumulation, or abnormal vasculature—supporting the models’ interpretability and their potential for clinical trust.Results show that classical architectures like VGG16 and VGG19 outperform some modern lightweight models, with precision and recall rates exceeding 96%. These findings suggest that well-tuned, relatively shallow networks can efficiently distinguish retinal pathologies. The study concludes that transfer learning with pre-trained models is a promising approach for early and automated diagnosis of retinal diseases, potentially integrating into clinical workflows to enable timely interventions, improve patient outcomes, and increase access to accurate retinal assessments worldwide.