HYBRID EFFICIENTNET-TRANSFORMER ARCHITECTURE FOR AUTOMATED DETECTION OF GLAUCOMA

In recent years, the early and accurate diagnosis of Glaucoma, a leading cause of permanent blindness, has highlighted the importance of effective treatment and better patient care. Conventional deep learning methods that only use Convolutional Neural Networks (CNNs) have shown significant performance in classification tasks involving fundus images, but they often fail to capture global contextual dependencies that are crucial for fine-grained disease categorization. Addressing this limitation, a Hybrid CNN-Transformer architecture has been proposed that uses the local feature extraction capability of a pretrained EfficientNetB0 backbone and augments it with the global modeling power of Transformer encoders. The EfficientNetB0 module extracts robust spatial features, which are improved using Transformer layers to model inter-feature relationships throughout the image. The proposed Hybrid EfficientNet-Transformer model was trained and validated on a dataset created by combining five datasets containing Fundus images, employing standard preprocessing pipelines and label encoding strategies. Extensive experiments show that the proposed hybrid approach attains a 99% high classification accuracy on the unseen test set, outperforming many existing HybridCNNs and transfer learning models. These results indicate that the fusion of CNNs and Transformers offers a powerful framework for high-precision glaucoma detection, creating new prospects for the development of automated ophthalmic screening tools.