SF Net: A Pyramid-Based Feature Fusion Convolutional Neural Network With Embedded Squeeze-and-Excitation Mechanism for Retinal OCT Image Classification

Abstract

Age-related macular degeneration (AMD) and diabetic macular edema (DME) are among the leading causes of blindness worldwide, and optical coherence tomography (OCT) analysis plays a crucial role in diagnosing and treating ocular diseases. While deep learning has been extensively applied to OCT image classification, existing methods often require large-scale training datasets. However, the inherent challenges of medical image acquisition make large datasets difficult to obtain. Therefore, it is desirable to develop models that can achieve high performance even with limited training data. Moreover, most current approaches rely solely on features extracted from the final network layer, whereas incorporating intermediate feature maps can further enhance classification accuracy. In this study, a novel end-to-end multi-scale classification framework, termed SF Net (squeeze-and-excitation (S) embedded feature fusion pyramid (F) convolutional neural network), is proposed for the reliable diagnosis of eye conditions, including normal retinal images and three clinical categories: early and late stages of age-related macular degeneration (AMD) and diabetic macular edema (DME). The effectiveness of the proposed method is evaluated on two datasets: a national dataset collected at Noor Eye Hospital (NEH) and a publicly available dataset from the University of California, San Diego (UCSD). The experimental results demonstrate that the proposed multi-scale method outperforms all well-known OCT classification frameworks. Despite a significant reduction in the training dataset size, the model's performance still exceeds that of most comparable networks.