A Multi-branch and Attention Based CNN Architecture for the Classification of Retinal Diseases from OCT Images

Retinal diseases are becoming increasingly prevalent, and their accurate diagnosis demands precise and reliable identification techniques. Although numerous deep learning-based methods have been developed to classify these diseases using OCT images, many of them fail to focus on extracting the most discriminative features for optimal classification. Additionally, several previous studies have solely relied on binary classification, rather than multi-class methods. To address these challenges, this study presented a multi-branch CNN architecture with an attention mechanism, capable of identifying the most informative features for better classification. The upper part of the architecture was designed with depth to extract increasingly complex and abstract features, while the shallow lower part was intended to remove noise and artifacts and extract different types of features. The primary contribution of this study was the incorporation of the convolutional block attention module (CBAM) in the CNN architecture. By dynamically weighing the feature maps of the CNN, CBAM allows the network to concentrate on the most informative regions of an image while discarding irrelevant features. The proposed CNN architecture remarkably outperformed most recent competing investigations on OCT images, achieving 96% accuracy and 96% f1-score.