DCViT: A Hybrid Dilated Convolution-Vision Transformer Model for Retinal Disease Classification Using OCT Images
IEEE Access · 2026 · 14 · 107650-107661
DOI: 10.1109/access.2026.3713423 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Zahra Davarzani und 1 weiteren Autor:innen. Die Arbeit nennt 47 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Retinal disorders are a significant cause of both blindness and temporary or permanent visual impairment. Early diagnosis plays a crucial role in effective treatment. This study proposes a hybrid model that leverages Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) to classify retinal diseases into four categories: Choroidal Neovascularization (CNV), Diabetic Macular Edema (DME), Drusen, and normal, using Optical Coherence Tomography (OCT) images. CNNs have demonstrated remarkable success in extracting local features and have significant contributions to medical diagnosis. However, accurately classifying retinal diseases requires understanding the relationship between localized lesions and the global retinal structure. With their self-attention mechanism, transformers enable the extraction of long-term dependencies in medical images, which is vital for precise classification. Existing approaches often fail to explicitly model the interaction between lesion-scale features and the global retinal structure in OCT images. Unlike generic CNN-transformer hybrids, the proposed architecture is specifically designed to encode multi-scale retinal lesion patterns prior to global attention modeling. This research proposes a hybrid Dilated Convolution-Vision Transformer (DCViT) model incorporating a Dilated Convolutional Module (DCM) to extract feature maps from OCT images. This module uses a dilated convolution kernel with a variable dilation rate, ranging from 1 to 3, to capture multi-scale features. The derived feature maps are passed through a pre-trained ViT, which enhances the ability to capture long-range dependencies. Finally, the features are classified into the four retinal disease categories. This innovative integration of DCM with ViT not only enhances feature extraction but also shows promising potential for automated retinal OCT image analysis. The DCViT model was trained, evaluated, and tested on the OCT2017 dataset, achieving an accuracy of 99.69% and demonstrating competitive performance among recent state-of-the-art methods. The proposed model, with its high generalizability, may contribute to enhancing medical diagnostics and patient care.
Zitieren
APA
Davarzani, Z., & Mosavi, M. R. (2026). DCViT: A Hybrid Dilated Convolution-Vision Transformer Model for Retinal Disease Classification Using OCT Images. IEEE Access, 14, 107650-107661. https://doi.org/10.1109/access.2026.3713423
Vancouver
Davarzani Z, Mosavi MR. DCViT: A Hybrid Dilated Convolution-Vision Transformer Model for Retinal Disease Classification Using OCT Images. IEEE Access. 2026;14:107650-107661. doi:10.1109/access.2026.3713423
BibTeX
@article{Davarzani2026,
author = {Davarzani, Zahra and Mosavi, Mohammad Reza},
title = {DCViT: A Hybrid Dilated Convolution-Vision Transformer Model for Retinal Disease Classification Using OCT Images},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {107650--107661},
doi = {10.1109/access.2026.3713423}
}
EUVIMED