22 May 2025 in IEEE Access
Automatic Diabetic Retinopathy classification is crucial as it aids in determining the disease stage, enabling timely treatment to prevent progression to low vision and significantly reducing the diagnostic burden on ophthalmologists. Efficient and accurate network models are essential for clinical applications. In this paper, we propose a deep learning method for DR diagnosis and grading based on multi-scale feature fusion and attention guidance. This approach employs dilated convolutions with varying dilation rates to expand the receptive field and integrate features at multiple scales. Additionally, an attention mechanism based on spatial directional feature aggregation is embedded in the inverted residual modules to guide the model to focus on critical lesion features in fundus images. Furthermore, transformers are utilized to extract global features, enhancing model performance. Experiments conducted on the APTOS-2019 and DDR datasets demonstrate that our method achieves a diagnostic accuracy of 99.2% and grading accuracy of 88.6% on the APTOS-2019 dataset, and a grading accuracy of 84.8% on the DDR dataset. Experimental results show that the proposed method performs excellently and outperforms many recently proposed popular DR diagnosis and classification methods.
19 August 2025 in IEEE Access
Diabetic retinopathy (DR) is a leading cause of preventable blindness worldwide. Early and accurate diagnosis is essential to prevent irreversible vision loss. However, current manual grading methods are inefficient, underscoring the need for automated computer-aided diagnosis (CAD) systems. Although deep learning has advanced DR grading, most approaches focus solely on spatial features, overlooking complementary frequency-domain information and facing challenges in detecting subtle lesions and modeling inter-lesion dependencies. This study addresses these limitations by proposing a novel architecture, the Joint Frequency-Spatial Perception Vision Transformer (JFSP-ViT), which integrates spatial and frequency-domain learning to improve DR severity assessment. JFSP-ViT employs a hierarchical design with Inverted Residual Blocks (IRB) and Inverted Residual Attention Blocks (IRAB) to enhance spatial feature representation progressively. Furthermore, the Dual-Domain Perception Transformer (DDPT) module enables joint-domain representation learning through spatial and frequency self-attention mechanisms, while a cross-domain attention fusion module facilitates interaction between the two domains. Experimental results on public DR datasets demonstrate that JFSP-ViT achieves superior performance in multi-level DR grading, surpassing state-of-the-art methods in both accuracy and robustness, particularly in detecting subtle lesions and modeling inter-lesion dependencies. These findings emphasize the significance of dual-domain learning in automated DR diagnosis and suggest that JFSP-ViT can serve as a reliable tool for early and precise DR assessment.
1 January 2026 in IEEE Access
1 January 2026 in IEEE Transactions on Multimedia