SCAN-FIQA: Multi-scale cross-attention network for fundus image quality assessment

Currently, the performance and reliability of AI-driven fundus image diagnostic models are critically constrained by the quality of real-world data. However, in clinical scenarios, the quality of fundus images often varies significantly due to factors such as differences in imaging equipment, suboptimal acquisition conditions, and inter-patient variability, etc. The presence of low-quality data may introduce bias during model training, compromising the generalizability of diagnostic models in practical applications. Although deep learning has shown promise in fundus image quality assessment (FIQA), current methods struggle to robustly handle challenges such as pathological variability, structural ambiguity, and imaging artifacts. To address these limitations, we propose SCAN-FIQA, a novel Three-Branch Multi-Scale Cross-Attention Network that integrates three complementary modules to achieve robust and fine-grained FIQA. Firstly, the Wavelet-Based Denoising Module (WBDM) suppresses high-frequency noise while preserving structural details. Secondly, the Local Branch (LB) captures local detailed information for fine-grained quality assessment. Thirdly, the Boundary-Aware Multi-Scale Network (BaMS-Net) captures multi-scale global semantics and local details through two parallel and independent feature extraction branches with coarse and fine granularity. By jointly leveraging denoising, boundary-awareness, and hierarchical feature fusion, SCAN-FIQA effectively mitigates quality-induced challenges in fundus imaging. Extensive experiments conducted on two FIQA datasets, including the public EyeQ dataset and the private SY-FIQA dataset, demonstrate that SCAN-FIQA achieves outstanding performance. This framework offers a reliable solution for filtering low-quality images, thereby enhancing diagnostic accuracy and supporting robust clinical decision-making in real-world settings.