Retinopathy of prematurity (ROP) is a major cause of vision loss and blindness among premature infants. Timely screening, diagnosis, and intervention can effectively prevent the deterioration of ROP. However, there are several challenges in ROP diagnosis globally, including high subjectivity, low screening efficiency, regional disparities in screening coverage, and severe shortage of pediatric ophthalmologists. The application of artificial intelligence (AI) as an assistive tool for diagnosis or an automated method for ROP diagnosis can improve the efficiency and objectivity of ROP diagnosis, expand screening coverage, and enable automated screening and quantified diagnostic results. In the global environment that emphasizes the development and application of medical imaging AI, developing more accurate diagnostic networks, exploring more effective AI-assisted diagnosis methods, and enhancing the interpretability of AI-assisted diagnosis, can accelerate the improvement of AI policies of ROP and the implementation of AI products, promoting the development of ROP diagnosis and treatment.
ObjectivesTo develop a fundus photography (FP) image lesion recognition model based on the EfficientNet lightweight convolutional neural network architecture, and to preliminary evaluate its recognition performance. MethodsA diagnostic test. The data was collected in the Department of Ophthalmology at Sichuan Provincial People's Hospital from June 2023 to June 2025. A lightweight 16-category lesion recognition model was constructed based on deep learning and 610 072 FP images. The FP images were sourced from Sichuan Provincial People's Hospital as well as the APTOS, Diabetic Retinopathy_2015, Diabetic Retinopathy_2019, and Retinal Disease datasets. Model performance was evaluated as follows: first, testing was performed on four independent external validation sets using metrics such as accuracy, F1 score (the harmonic mean of precision and recall), and the area under the receiver operating characteristic curve (AUC) to measure the model's generalizability and accuracy. Second, the classification results of the model were compared with those of junior and mid-level ophthalmologists (two each) using the overlapping confidence interval (CI) comparison method to assess the clinical experience level corresponding to the model's medical proficiency. ResultsThe model achieved an accuracy of 96.78% (59 039/61 003), an F1 score of 82.51% (50 334/61 003), and an AUC of 99.93% (60 960/61 003) on the validation set. On the four external validation sets, it achieved an average accuracy of 87.77% (57 358/65 350), an average precision of 87.06% (56 894/65 350), and an average Kappa value of 82.28%. The average accuracy of FP image lesion identification for junior and mid-level ophthalmologists was 79.00% (79/100) (95%CI 67.71-90.29) and 87.00% (87/100) (95%CI 77.68-96.32), respectively. ConclusionsA 16-category FP image lesion recognition model is successfully constructed based on the EfficientNet lightweight convolutional neural network architecture. Its clinical performance preliminarily reaches the level of mid-level ophthalmologists.
