Abstract
INTRODUCTION: Cataracts, the leading cause of reversible blindness globally, require timely detection and intervention for effective prevention of blindness. Artificial intelligence can assist in massive screening, however, existing models often trained on homogeneous, single-center data, suer from poor generalizability. METHODS: To address this challenge, we developed and validated a deep learning model trained on a large-scale, multicenter, real-world dataset comprising 22,094 slit-lamp images from 21 ophthalmic institutions across 12 provinces and municipalities in China. We designed a cascaded framework that emulates the sequential reasoning of a clinical diagnostic workflow, a methodological approach for ensuring reliability on noisy, real-world data. It first performs an automated quality assessment, then screens for common confounders like pterygium, and finally conducts a differential diagnosis among cataract, post-cataract surgery, other ocular diseases, and healthy eyes. Within this framework, we evaluated several deep learning architectures. RESULTS: In the cataract classification task, the leading models demonstrated excellent performance on an independent test set. For instance, the ResNet50-IBN based model achieved an accuracy of 93.74%, specificity of 97.74% and an area under the curve (AUC) of 95.30%. DISCUSSION: This study demonstrates that training on multicenter, real-world data yields a robust and generalizable model, providing a powerful tool for largescale ophthalmic screening. Specifically, our model establishes a methodological blueprint for developing trustworthy medical deep learning systems.