FANet: Joint Spatial–Frequency Domain Framework for Retinal Disease Classification From OCT Images

The increasing prevalence of retinal diseases fuels the demand for automated diagnostic technologies. However, the sparsity and multiscale features of retinal optical coherence tomography (OCT) images, as well as the high demands on the applicability of detection algorithms in practical settings, pose great challenges to the accurate and robust retinal OCT image classification algorithms. To address these issues, this article introduces a frequency-aware network (FANet) for retinal OCT-based disease classification, leveraging complementary spatial- and frequency-domain information. FANet leverages a multiscale Laplacian pyramid (MSLP) to capture fine-grained details in the spatial domain and a harmonic feature enhancement (HFE) module to model broader contextual information in the frequency domain. These diverse features are integrated through a feature fusion unit (FFU) to form a comprehensive feature representation. Our experiments demonstrate that FANet outperforms existing state-of-the-art (SOTA) models, achieving a 2.3% accuracy improvement on our retinal OCT dataset while maintaining low computational complexity. Furthermore, the extended experiments comprehensively demonstrate FANet’s robustness to noise interference, strong generalization capability across different datasets, adaptability to imbalanced multiclass classification scenarios, and sensitivity to varying amounts of training data.