Deep Learning-Based Retinal Disease Classification Using OCT Images: A Comparative Evaluation of Twenty Models
IEEE Access · 2026 · 14 · 74958-74968
DOI: 10.1109/access.2026.3693917 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Nader A. Rahman Mohamed. Die Arbeit nennt 31 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0). Gefördert von: Prince Sattam bin Abdulaziz University.
Abstract (Originalsprache)
Retinal diseases represent a major cause of visual impairment worldwide, with conditions such as age-related macular degeneration, diabetic macular edema, retinal vein occlusion, and retinal artery occlusion affecting large patient populations. Delayed diagnosis increases the risk of permanent vision loss. Optical coherence tomography supports early detection through high resolution visualization of retinal structures, yet interpretation requires expert clinicians and increases workload in large screening programs. Automated analysis systems support efficient clinical decision making. This study evaluates twenty different deep learning models, including traditional convolutional neural network and nineteen pretrained models for automated classification of retinal diseases using OCT images. The dataset includes 3,678 clinically verified scans collected from Prince Sattam bin Abdulaziz University Hospital and covers ten retinal disease categories. Two experienced retinal specialists performed independent labeling and verification. Transfer learning supported adaptation of pretrained models to the clinical dataset. Performance evaluation used accuracy, precision, recall (sensitivity), and F1 score. Results show clear variation across architectures. DarkNet53 achieved the highest performance with 95.93 percent accuracy and an F1 score of 0.955. ResNet50 and DarkNet19 followed with accuracy above 94 percent. DenseNet201 and VGG models also demonstrated strong classification capability. Lightweight models such as MobileNetV2 and ShuffleNet achieved moderate performance with reduced computational cost. Misclassification occurred mainly between clinically similar conditions, including Age-Related Macular Degeneration with Macular Neovascularization and with suspected Macular Neovascularization, as well as retinal vein occlusion and diabetic macular edema. Statistical analysis using ANOVA and Friedman ranking confirmed significant differences between models and identified DarkNet53 as the most consistent architecture. The findings support the use of deep learning models for automated OCT interpretation and highlight suitable architectures for clinical retinal disease screening systems.
Förderung
- Prince Sattam bin Abdulaziz University
Zitieren
APA
Mohamed, N. A. R. (2026). Deep Learning-Based Retinal Disease Classification Using OCT Images: A Comparative Evaluation of Twenty Models. IEEE Access, 14, 74958-74968. https://doi.org/10.1109/access.2026.3693917
Vancouver
Mohamed NAR. Deep Learning-Based Retinal Disease Classification Using OCT Images: A Comparative Evaluation of Twenty Models. IEEE Access. 2026;14:74958-74968. doi:10.1109/access.2026.3693917
BibTeX
@article{Mohamed2026,
author = {Mohamed, Nader A. Rahman},
title = {Deep Learning-Based Retinal Disease Classification Using OCT Images: A Comparative Evaluation of Twenty Models},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {74958--74968},
doi = {10.1109/access.2026.3693917}
}
EUVIMED