Toward Efficient Identification of Retinal Diseases: A Lightweight Convolutional Neural Network-Based Approach Using Optical Coherence Tomography.

Retinal diseases are a major cause of both temporary and permanent vision loss, making early detection and treatment essential to prevent irreversible damage, as the retina's ability to transmit light signals to the brain can be compromised. While deep learning has shown strong potential in diagnosing such conditions using optical coherence tomography (OCT) images, many existing models are highly complex and computationally intensive, making them impractical for deployment on edge devices in clinical settings. To address this issue, we propose a diagnostic framework based on a lightweight deep-learning architecture specifically designed for efficient retinal disease detection. Our model integrates two key components: a lite convolution block, utilizing depthwise separable convolutions for computational efficiency, and a global-local fusion block, which captures both fine-grained local and contextual global features. A squeeze-and-excitation mechanism further refines channel-wise feature importance, all while keeping the parameter count to only 0.27 million. We evaluate our model on three benchmark datasets-OCT 2017, OCT C8 and OCTDL- achieving accuracies of 99.70%, 95.00% and 97.26%, respectively. Our approach demonstrates strong and stable performance validated by confusion matrix and ROC analysis, while Grad-CAM visualizations enhance interpretability, collectively aiming to offer a practical and efficient solution for real-time retinal disease diagnosis.