Illumination-enhancement guided CNN-Transformer for robust grading of diabetic retinopathy from fundus images
Results in Optics · 2026-03 · 24 · 101026
DOI: 10.1016/j.rio.2026.101026 ↗Kurz gefasst
Fachartikel in der Zeitschrift Results in Optics, 2026-03. Verfasst von Hrijuta Datta und 1 weiteren Autor:innen. Die Arbeit nennt 34 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0).
Abstract (Originalsprache)
Diabetic Retinopathy (DR) is a deteriorative retinal disease with the risk of irreversible vision impairment, resulting from the severe complications of a chronic disorder known as Diabetes Mellitus (DM). Prevention and precaution at an early stage, with accurate detection of DR severity level, are significantly effective for on-time clinical intervention and vision preservation. This paper describes an enhancement-based ensemble classifier deep learning framework for automated DR grading using retinal fundus images. The proposed work integrates a two-stage methodology comprising enhancement and classification. In the first stage, the retinal images from three public datasets Asia Pacific Tele-Ophthalmology Society (APTOS) 2019 Blindness Detection, Diabetic Retinopathy Dataset (DDR), and Indian Diabetic Retinopathy Image Dataset (IDRiD) and a private dataset from PC Chatterjee Memorial Eye Hospital undergo standardized preprocessing operations such as cropping, normalization, and Gaussian blur-based noise removal, followed by lesion localization via color space separation. The preprocessed images are then enhanced using Zero-Reference Deep Curve Estimation++ (Zero-DCE++) integrated with dynamic Contrast Limited Adaptive Histogram Equalization (CLAHE), which amplifies low-contrast and lesion regions for improved feature extraction. In the second stage, an ensemble of six advanced architectures, Convolutional Network Next (ConvNeXt), Residual Network with 152 Layers (ResNet152), Residual Network with 50 Layers (ResNet50), Densely Connected Convolutional Network with 121 Layers (DenseNet121), Extreme Inception (Xception), and Shifted Window Transformer (Swin Transformer), is utilized for DR grading. The proposed model achieved classification accuracies of 98.54% on the APTOS 2019 dataset,98.14% on the DDR, and 97.29% on IDRiD, highlighting the significance of the proposed model, its reliability, and effectiveness as a clinical decision support system for DR screening and severity assessment.
Zitieren
APA
Datta, H., & Bhandari, A. K. (2026). Illumination-enhancement guided CNN-Transformer for robust grading of diabetic retinopathy from fundus images. Results in Optics, 24, 101026. https://doi.org/10.1016/j.rio.2026.101026
Vancouver
Datta H, Bhandari AK. Illumination-enhancement guided CNN-Transformer for robust grading of diabetic retinopathy from fundus images. Results in Optics. 2026;24:101026. doi:10.1016/j.rio.2026.101026
BibTeX
@article{Datta2026,
author = {Datta, Hrijuta and Bhandari, Ashish Kumar},
title = {Illumination-enhancement guided CNN-Transformer for robust grading of diabetic retinopathy from fundus images},
journal = {Results in Optics},
year = {2026},
volume = {24},
pages = {101026},
doi = {10.1016/j.rio.2026.101026}
}
EUVIMED