FReCSA-ONFELNet: a novel deep learning architecture for multiclass classification of retinal pathologies

Abstract

The high prevalence of ophthalmic diseases has become a major threat to visual health. Traditional diagnosis relies heavily on manual assessment, which is time-consuming and prone to subjective variability, highlighting the urgent need for efficient and reliable automated methods. This study proposes a novel deep learning framework, FReCSA-ONFELNet, designed for the automated classification of diabetic retinopathy, glaucoma, cataract, and normal retina. The framework integrates three innovative modules: the Frequency-Regulated Channel–Spatial Attention (FReCSA), the Optic Nerve Feature Enhancement Layer (ONFEL), and the Multi-Scale Feature Fusion (MSFF). Specifically, FReCSA leverages frequency-domain information to enhance channel attention discriminability and accurately localize lesion regions; ONFEL strengthens the representation of key ophthalmic structures through a multi-scale convolutional design; MSFF captures both local and global lesion features across multiple scales. Additionally, a structural consistency loss function is introduced to effectively reduce false positives in pathological categories. Experiments conducted on a dataset of 4,180 fundus images demonstrate that the proposed model achieves an overall classification accuracy of 94%, and significantly outperforms advanced architectures such as VGG16, VGG19, AlexNet, ResNet, and InceptionNet-v4 in terms of precision, recall, and F1-score. Overall, FReCSA-ONFELNet exhibits advantages in both performance and robustness, indicating strong potential for application in early screening and clinical-assisted diagnosis of retinal diseases.