RetinoNet: A Hybrid CNN with Ghost Blocks and Dual Attention for Accurate Multi-Class Diabetic Retinopathy Classification
Year: 2026
Type: conference-paper
Abstract: Diabetic retinopathy (DR) is a leading cause of vision loss among individuals with diabetes, necessitating timely and accurate screening to prevent irreversible damage. While deep learning approaches have demonstrated promise in automated DR detection, challenges including class imbalance, difficulty in detecting subtle lesions, and high computational demands limit their effective deployment in real-world applications. To address these limitations, we propose RetinoNet, a hybrid convolutional neural network that integrates standard convolutional layers with Ghost Blocks and dual attention mechanisms (Squeeze-and-Excitation and Convolutional Block Attention Module) to efficiently capture both global contextual patterns and localized retinal abnormalities. RetinoNet is designed for five-class DR classification, achieving robust performance across all severity levels while maintaining computational efficiency and demonstrating potential for deployment in resource-constrained clinical environments. Comprehensive evaluations of the APTOS 2019 dataset show that RetinoNet demonstrates competitive performance compared to several baseline CNN architectures, achieving an overall accuracy of 95.00%, with strong precision, recall, and F1-score performance across all classes. Ablation studies further highlight the strength of the proposed architecture components in enhancing feature discrimination and generalization. These results suggest that RetinoNet provides a reliable and scalable framework for automated diabetic retinopathy screening and diagnosis.
Language: English
Topic: Retinal Imaging and Analysis
Subfield: Radiology, Nuclear Medicine and Imaging
Field: Medicine
Domain: Health Sciences
Open Access status: hybrid
OpenAlex ID: W7204467165