APHF-VIS: Advanced Privacy-Preserving Hybrid Framework with Vision and Swin Transformer for retinal disease detection

This is the authors' abstract. We don't add key points for this paper.

: Retinal disease detection plays a major role in preserving vision loss and providing timely clinical intervention. Although, existing automated diagnostic systems face several challenges such as multi-modal feature redundancy, loss of hierarchical vascular relationships and attention misalignment, which significantly degrade clinical interpretability and classification accuracy. Hence, a novel Advanced Privacy-Preserving Hybrid Transformer Framework is designed to improve stabilize attention, feature representation and prevent vascular structure. The framework captures global anatomical context to decrease misclassification of widely distributed lesions while simultaneously modeling local hierarchical patterns to enhance the detection of fine-grained abnormalities such as hemorrhages and microaneurysms. Lesion textures and subtle vessel are further emphasized to tackle redundant representations and enhance discrimination. Feature fusion and optimization ensure complementary information integration, while attention stabilization improves robustness across heterogeneous clinical environments. In addition, a privacy-preserving collaborative learning strategy enables multi-institutional training without exposing sensitive medical data. The proposed framework efficiently combines global, textural and local information, preserves hierarchical vascular structures, ensures lesion-focused attention and offers interpretable outcomes, thus making it suitable for reliable and scalable retinal disease detection.