Using Deep Learning to Classify Retinal Diseases from OCT Images
Web3 Journal: ML in Health Science · 22.06.2026 · 3 (1) · d22062026
DOI: 10.62487/1t999t50 ↗Kurz gefasst
Fachartikel in der Zeitschrift Web3 Journal: ML in Health Science, online erschienen am 22.06.2026. Verfasst von Foma Molchanov und 1 weiteren Autor:innen. Die Arbeit nennt 16 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-SA 4.0).
Abstract (Originalsprache)
Automated analysis of optical coherence tomography (OCT) images has become increasingly important for the early detection of retinal diseases. In this study, deep learning models were developed for multi-class retinal disease classification using the OCT2017 dataset, which includes four categories: choroidal neovascularization (CNV), diabetic macular edema (DME), DRUSEN, and NORMAL retinal images. Two architectures were evaluated: DenseNet201 and an attention-based ensemble model combining EfficientNet-B3 and ResNet50. Transfer learning with ImageNet-pretrained weights was used together with a standardized preprocessing pipeline that included image resizing, grayscale-to-three-channel conversion, normalization, and online data augmentation. To improve performance on underrepresented classes, class-weighted focal loss was applied during training, particularly to enhance DRUSEN detection. Model evaluation was performed using five-fold stratified cross-validation. Among the tested models, DenseNet201 achieved the best overall performance, with a mean cross-validation accuracy of 96.69% ± 0.10%, macro F1-score of 0.9531, macro ROC-AUC of 0.9949, Cohen’s κ of 0.9510, and Expected Calibration Error (ECE) of 0.0112. Class-wise analysis showed excellent performance for CNV, DME, and NORMAL categories, while DRUSEN remained comparatively more challenging because of its subtle visual characteristics. Nevertheless, DRUSEN recall improved to 91.7% after the introduction of class-weighted focal loss. Calibration analysis further demonstrated that the proposed model produced reliable probability estimates suitable for clinical decision-support applications. Overall, the results show that DenseNet201 provides a reliable and clinically useful approach for automated retinal disease classification using OCT images.
Zitieren
APA
Molchanov, F., & Pal, S. (2026). Using Deep Learning to Classify Retinal Diseases from OCT Images. Web3 Journal: ML in Health Science, 3(1), d22062026. https://doi.org/10.62487/1t999t50
Vancouver
Molchanov F, Pal S. Using Deep Learning to Classify Retinal Diseases from OCT Images. Web3 Journal: ML in Health Science. 2026;3(1):d22062026. doi:10.62487/1t999t50
BibTeX
@article{Molchanov2026,
author = {Molchanov, Foma and Pal, Sukhwant},
title = {Using Deep Learning to Classify Retinal Diseases from OCT Images},
journal = {Web3 Journal: ML in Health Science},
year = {2026},
volume = {3},
number = {1},
pages = {d22062026},
doi = {10.62487/1t999t50}
}
EUVIMED