Lightweight deep learning approach for retinal OCT image classification: A CNN with hybrid pooling and optimized learning

Abstract

Optical coherence tomography (OCT) is a non-invasive technique through which a retina specialist can see the structure behind the eye. This technol ogy offers a key role to identify various abnormalities in the retina: Drusen, diabetic macular edema (DME) and choroidal neovascularization (CNV). However, manual analysis of OCT scans can be time-consuming and prone to variability among clinicians. To address this challenge, we present a lightweight and explainable deep learning-based approach for automatic classification of retinal OCT images. The primary goal of this research is a model that delivers high diagnostic accuracy. A computer-aided suggestive method can help retinal doctors automatically classify the anomalies with more confidence and precision. In this paper, we proposed a novel approach based on deep learning: a six-layer convolutional neural network (CNN) integrated with hybrid pooling for effective feature extraction. Data augmentation and exponential learning rate is implemented to handle data imbalance between classes and for stabilized learning consecutively. Our proposed approach achieved 98.75% of accuracy while testing on the dataset. To further enhance the interpretability of the model, we also integrate explainable AI (XAI) using class activation mapping (CAM) to visualize the critical regions in the retina that contribute to the classification decisions.