F2D-RetinaNet: a feature-to-decision dual-path network with fractional operators for retinal disease detection and grading

Abstract

Visible pathological abnormalities and vascular integrity help diagnose retinal diseases. This paper introduces a unified feature-to-decision framework that first extracts vasculature and enhances lesions. First, this study proposes the fractional anisotropic diffusion–based hybrid enhancement (FADHE) method that selectively amplifies lesion contrast and vessel visibility in the green channel using long-memory fractional diffusion dynamics. Second, this study introduces a fractional Hessian-based attention U-Net (FHAU-Net), a multi-scale vessel segmentation model that leverages fractional Hessian responses across eight fractional orders and attention-guided skip connections to capture fine vessel structures. Finally, this work designs a dual-branch disease grading model, F2D-RetinaNet, which jointly processes the enhanced image and vessel map to learn lesion-related and vascular feature representations for the recognition and grading of retinal diseases. The FHAU-Net model achieved a Dice coefficient of 0.8987, a sensitivity of 0.9084, an accuracy of 0.9804, and a specificity of 0.9876 across the unified test set, which contains variations in resolution, acquisition site, camera type, and imaging conditions. The F2D-RetinaNet is evaluated for Diabetic retinopathy (DR) detection, DR grading, and multiclass retinal disease classification tasks. Extensive testing and comparisons to the baseline procedures support the reliability of the proposed methodology on the datasets evaluated here.