A Novel Transformer-CNN Hybrid Deep Learning Architecture for Robust Broad-Coverage Diagnosis of Eye Diseases on Color Fundus Images

We present a novel and robust deep-learning architecture that takes into account the pathological characteristics of eye diseases on color fundus images. The proposed hybrid architecture is capable of accurately recognizing the local features on the one hand and detecting the peripheral abnormalities of the specific eye disease on the other, and diagnosing the respective eye disease on this basis. The hybrid architecture provides robust local feature extraction and enhances global contextual understanding via hierarchical self-attention mechanisms. In contrast to previous models, which were limited to binary or poorly differentiated classification tasks, our approach considers the most clinically relevant eye diseases (here 9 diseases), including central serous chorioretinopathy (CSCR), diabetic retinopathy (DR), disc edema, glaucoma, macular scar, myopia, pterygium, retinal detachment, and retinitis pigmentosa (RP). The architecture was validated using a peer-reviewed real-world dataset and stratified 5-fold cross-validation. It achieved an average balanced accuracy of 81.91%, an average accuracy and True Positive Rate (TPR) of 76.40%, and a True Negative Rate (TNR) of 95.60% across all classes. For individual diseases, the model reached an TPR of 88.93% for DR, 89.93% for RP, and 61.23% for Glaucoma. The results show that the architecture is powerful and can distinguish between the 10 different classes (9 disease classes + healthy class). This makes it particularly suitable for pre-screening applications in telemedicine, for screening programs at the population level, and for integration into clinical decision support systems.