FF-ResNet-DR model: a deep learning model for diabetic retinopathy grading by frequency domain attention

Diabetic retinopathy (DR) is a major cause of vision loss. Accurate grading of DR is critical to ensure timely and appropriate intervention. DR progression is primarily characterized by the presence of biomarkers including microaneurysms, hemorrhages, and exudates. These markers are small, scattered, and challenging to detect. To improve DR grading accuracy, we propose FF-ResNet-DR, a deep learning model that leverages frequency domain attention. Traditional attention mechanisms excel at capturing spatial-domain features but neglect valuable frequency domain information. Our model incorporates frequency channel attention modules (FCAM) and frequency spatial attention modules (FSAM). FCAM refines feature representation by fusing frequency and channel information. FSAM enhances the model's sensitivity to fine-grained texture details. Extensive experiments on multiple public datasets demonstrate the superior performance of FF-ResNet-DR compared to state-of-the-art models. It achieves an AUC of 98.1% on the Messidor binary classification task and a joint accuracy of 64.1% on the IDRiD grading task. These results highlight the potential of FF-ResNet-DR as a valuable tool for the clinical diagnosis and management of DR.