Seg-Swin: A Dual-Attention Transformer Model for Advanced AMD Classification and Lesion Detection Using Color Fundus Imaging

Age-related macular degeneration (AMD) is a prevalent retinal disorder in the elderly, often leading to significant vision impairment. The diagnosis of AMD is confirmed through various medical imaging modalities, with color fundus photography (CFP) being a primary tool. The detection and staging of AMD severity depend on several factors, including the number and size of drusen, the presence of pigmentary changes, geographic atrophy, and neovascularization, all of which are identifiable through CFP. In this study, we introduce an innovative dual-vision transformer-based network designed to automatically detect AMD and classify its severity into either dry AMD or wet AMD using CFP. Early diagnosis and accurate staging of AMD are crucial in mitigating the progression of the disease, making this work particularly valuable. Our proposed model, Seg-Swin, leverages a dual attention-based transformer network architecture, comprising two key stages. The first stage employs the SegFormer transformer model for the precise detection of AMD-related lesions, while the second stage utilizes the Swin transformer model to classify the detected lesions into dry or wet AMD. Our extensive experimental results demonstrate that the Seg-Swin model outperforms existing approaches, achieving remarkable diagnostic accuracy with metrics such as 98.7% accuracy, 99% sensitivity, 97.95% F1-score, and 98.24% specificity. By combining the strengths of advanced transformer models in both identification and classification tasks, the Seg-Swin model offers a comprehensive and powerful solution for detecting and staging AMD. The integration of these dual attention mechanisms allows the model to more precisely interpret complex retinal images, which is crucial for early diagnosis and accurate staging of AMD.