Attention-Driven Feature Fusion Integrating Swin Transformer and CNN Models for Improved Ocular Disease Classification

Fundus images are crucial for diagnosing ocular diseases, but timely and accurate classification remains chal-lenging due to complex retinal structures and significant class imbalances in datasets like the Ocular Disease Intelligent Recog-nition (ODIR). In this paper, we present a hybrid deep learning model that synergizes the Swin Transformer, Inception V3, and EfficientNetBO architectures to enhance ocular disease classi-fication. We improve multi-scale feature extraction by adding dilated convolutions to the Swin Transformer and integrating global average pooling into the Inception V3 and EfficientN etBO models. After these modifications, we unfreeze the last few layers of each model and fine-tune them on the ODIR dataset, which we partitioned into 80 % for training and 20 % for testing. We then concatenate the outputs and apply a self-attention mechanism to focus on the most informative features. To address the inherent class imbalance in the ODIR dataset, we use class-specific weights and employ focal loss during training, which alleviates the model's sensitivity to underrepresented classes and improves recall for these categories. Our model demonstrates superior performance across all eight ocular disease categories, achieving an average precision and recall of 90 %, with values ranging from 84% to 98%. Notably, it attains precision and recall of 98% for cataract detection and 96 % for myopia. These results underscore the effectiveness of combining convolutional neural networks and transformer architectures with attention-based feature fusion for comprehensive ocular disease classification.