<p>Glaucoma, as a chronic and progressive ocular disease, can cause vision impairment and blindness if ignored without suitable diagnosis and treatment, particularly in the early phases. The importance of early diagnosis of glaucoma is attributed to several reasons: it is one of the world’s most common causes of blindness, and the asymptomatic and subtle nature of its development. Recent developments in deep learning models have shown great opportunities for automating the detection of glaucoma using photographs of the retinal fundus. Still, the current convolutional neural networks (CNNs) depend on single-model schemes, which might not adequately capture the long-range spatial correlations and local fine-grained data required for reliable diagnosis. To address these limitations, this study proposes FusionNet-GD, a hybrid deep learning framework that integrates multi-stream CNN feature extraction with a Swin Transformer-based classification model. The proposed approach leverages the strengths of multiple CNN architectures: EfficientNetB0, ResNet50, and InceptionV3 to extract complementary features. Then, they are refined through principal component analysis (PCA) to reduce redundancy and computational complexity. These fused feature representations are subsequently processed by the Swin Transformer, which employs hierarchical self-attention mechanisms to effectively model both local and global contextual relationships in fundus images. The experiments were conducted using two benchmark datasets: LAG and EDC. FusionNet-GD achieves an accuracy of 99.07% on the LAG dataset and 93.54% on the EDC dataset compared to conventional CNN-based and CNN-Transformer hybrid models in terms of precision, recall, and F1-score. Hybrid deep learning with self-attention enables accurate early glaucoma detection using the FusionNet-GD model.</p>