SIGraph: Saliency Image-Graph Network for Retinal Disease Classification in Fundus Image

• DOI: 10.1609/aaai.v39i10.33090
• Publication Date: 2025-04-11T11:50:12Z
• Authors: Peng Zhang; Yuan Li; Haotian Song; Yankai Jiang; Yubo Tao; Hai Lin; Hongguang Cui
• Keywords: Fundus (uterus)

An efficient and precise diagnosis of retinal diseases is a fundamental goal for auxiliary diagnostic systems in ophthalmology. Inspired by the importance scattered subtle lesions manual disease diagnosis, recent research has achieved state-of-the-art performance mining information related to lesions, including their texture shape. However, spatial distribution patterns lesion areas, which are also crucial have been overlooked existing research. Neglecting these (e.g., ring microaneurysms diabetic macular edema) may negatively impact process. In this paper, we introduce Saliency-Image-Graph (SIGraph) network capture areas. We first employ saliency-based perception identify latent pixels. Subsequently, propose novel image-graph block efficiently global abundant pixels with minimal loss. By leveraging additional patterns, SIGraph achieves at least 1.5% gain across three datasets. Furthermore, ablation studies demonstrate that our can be integrated into other visual backbones effectively boost performance.