Glaucoma Detection from Fundus and OCT Images Using Attention-Based CNN and Transfer Learning Architectures

Glaucoma is a chronic eye disease that is leading cause of irreversible vision loss worldwide. Early and accurate classification of glaucoma is crucial for timely intervention and effective management. This study is undertaken to automate the classification of glaucoma using the dataset that contain fundus images and OCT images. The proposed work targets on deep learning-based architectures for glaucoma classification: CNN-based transfer learning models and CNN with an Attention mechanism. The transfer learning models leverage pretrained networks for efficient feature extraction, while the attention-based CNN enhances focus on glaucoma-specific regions. A 2D CNN combined with attention layers to enhances the capacity of the model to attend to the key areas of medical images. ResNet-18 is a type of deep learning model where every block in the network has a skip connection that adds the input back to the output. DenseNet121 is a deep convolutional neural network that improves feature reuse and gradient flow by introducing dense connections between layers. The models DenseNet121, ResNet18 and Attention based 2D CNN are trained separately on fundus and OCT data to classify whether an eye is affected by glaucoma. By comparing the accuracies and performance metrics of all the trained models, we determine which model and image type is more suitable for glaucoma classification. This helps in identifying the most effective imaging method for future AI-based diagnostic tools.