RetinoMamba-AnatomyQuery: Anatomy-Guided Query Learning for Efficient Vision Impairment Screening from Retinal Fundus Images

Automated analysis of fundus images is a promising technique for efficient, scalable retinal screening. In real-time screening, binary classification provides a straightforward first-line solution by differentiating normal and abnormal images, enabling prioritization of potentially abnormal cases for subsequent medical and disease-specific analysis. This study proposes RetinoMamba-AnatomyQuery, a lightweight model that integrates Vision Mamba-based global contextual representations, EfficientNet-B3-derived local features, anatomy–lesion query refinement, and anatomy-aware feature fusion. The Ocular Disease Intelligent Recognition (ODIR)-2019 dataset and Brazilian Multilabel Ophthalmological Dataset (BRSET) are harmonized to construct the internal cohort, yielding 10,842 eligible unique patients following expert-assisted quality control and data curation. The proposed model achieves accuracy of 97.93%, recall of 97.77%, specificity of 98.07%, precision of 97.87%, and an F1-score of 97.82%, outperforming state-of-the-art models. Independent evaluation on the external dataset (Retinal Fundus Multi-disease Image Dataset (RFMiD)) reports an accuracy of 96.21% and an F1-score of 95.99%, demonstrating sustained performance across an independent data distribution. Computational evaluation demonstrates a controlled footprint of 23.9 million parameters and 8.3 giga floating-point operations, supporting computationally efficient screening. Overall, the findings demonstrate the effectiveness of anatomy-conditioned global–local feature integration for reliable normal/abnormal fundus discrimination.