Dual branch fundus deep learning network as an enhanced multi classification system for ocular disease detection via hybrid feature fusion.

Ophthalmic diagnosis relies heavily on the interpretation of fundus images to identify a range of debilitating diseases. However, the presence of multiple, co-existing pathologies and the subtle visual cues associated with early-stage disease pose a significant challenge, necessitating the development of advanced diagnostic tools. We present DFD-Net (Dual-Branch Fundus Deep Learning Network), a deep learning solution that performs multi-label classification of various ocular diseases. The system achieves this by integrating the attribute maps from bilateral retinal photographs, like those from the right and left eyes. This study starts with a thorough pre-processing operation on the fundus images, represented in image rescaling, black border image cropping, contrast enhancement, and data augmentation of retinal photographs. Following this, the processing pipeline begins with a dual-branch feature extraction framework. A part of the system employs a ConvNeXt architecture to obtain complex semantic characteristics, while the parallel branch employs a U-Net encoder to capture fine-grained, structural characteristics across several scales. These complementary feature representations are then combined via a fusion mechanism. The fused feature representations are subsequently enhanced by a SENet Block for greater robustness through channel-wise recalibration. The resulting feature maps are reduced via Global Average Pooling, and then passed through a Dense layer for feature refinement. Finally, the features are classified by a Softmax output layer, consistent with the single-label multi-class formulation adopted in this work, to identify one of eight possible categories. The testing was implemented on the Ocular Disease Intelligent Recognition-Ophthalmic Image analysis (OIA-ODIR) dataset, which has retinal photographs that represent six distinct ophthalmic categories, macular degeneration, cataracts, glaucoma, hypertension, diabetic retinopathy, and myopia, images that present normal cases, and images that represent multiple other diseases, depicting eight different categories. The proposed DFD-Net system established precision, recall, F1- scores, and overall accuracy of 92.75%, 93.17%, 94.45%, and 93.77% on the On-Site testing collection and 91.50%, 93.19%, 92.11%, and 92.97% on the Off-Site testing collection, respectively. Ultimately, the suggested DFD-Net system proved high efficiency in precisely classifying multiple ocular diseases, offering a novel and outstanding approach for early detection of fundus diseases.