Development and validation of a blinding eye disease screening system based on lightweight convolutional neural networks: A diagnostic accuracy study.

To develop a cost-effective screening model for blinding eye diseases, assist primary medical institutions in improving screening efficiency, and reduce the risk of irreversible visual impairment in patients caused by delayed diagnosis. In this study, a lightweight ophthalmic disease screening model was constructed based on deep learning and 89,158 fundus photographs. Model performance evaluation was divided into 2 stages: first, testing was performed on 2 independent external validation datasets to assess the generalizability and stability of the model. Second, the classification performance of the artificial intelligence model was compared with that of ophthalmologists with different years of clinical experience to determine the clinical experience level corresponding to the model's diagnostic capability. The model achieved an average accuracy of 96.11%, F1-score of 82.39%, recall of 81.61%, receiver operating characteristic-area under the curve of 99.77%, precision of 96.44%, and specificity of 98.66% on the internal validation set; on external validation datasets, the average accuracy was 90.92%, average specificity 97.03%, and average precision 89.98%. In the comparative evaluation with clinicians, junior ophthalmologists achieved an accuracy of 74% (95% CI: 61.84-86.16%, α = 0.05) in identifying lesion-containing false-positive images, while mid-level ophthalmologists reached 88% (95% CI: 78.99-97.01%, α = 0.05). This study developed an artificial intelligence-based screening system for blinding eye diseases with performance comparable to that of mid-level ophthalmologists, providing an effective solution for promoting population-based screening of blinding eye diseases in primary care settings.