Classification of macular abnormalities using a lightweight CNN-SVM framework

Macular abnormalities are the main reason for central vision loss, especially in elderly people. Due to global population aging, a heavy burden will be placed on the health care system. Therefore, it is urgent and necessary to develop an automatic and intelligent tool to identify macular abnormalities. Optical coherence tomography is a non-invasive rapid imaging technique to diagnose macular abnormalities. We propose a lightweight convolutional neural network–support vector machine (CNN-SVM) framework consisting of a novel lightweight CNN backbone and an SVM classifier for the accurate detection of macular abnormalities. The CNN-SVM framework achieves excellent performance based on various metrics (precision, recall, F1-score, and accuracy) with an accuracy of 99.8% and demonstrates strong interpretability using heatmap visualization, outperforming several state-of-the-art models (Joint-Attention Network, OpticNet, MobileNet-V3, DenseNet-169, ResNet-50, lesion-aware CNN, Atten-ResNet, least-squares generative adversarial network and others). The proposed CNN-SVM framework is a feasible and reliable tool for the classification of macular abnormalities and shows potential for diagnostic ophthalmology in clinical practice.