Objective To establish a deep learning model by using ultra-wide-field (UWF) fundus photos of 3–12 year old children to take fundus photos and predict the axial length (AL) of children's eyes.

Methods Patients aged 3–12 years with refractive errors between September 2022 and December 2023 were retrospectively analyzed. Patients whose AL was measured and UWF images were taken within 1 month after measurement were selected as study subjects. A total of 177 UWF images from 100 patients were included. The patients were divided into four groups according to AL: (1) AL ≤ 22 mm; (2) 22 mm ＜ AL ≤ 23 mm; (3) 23 mm ＜ AL ≤ 24 mm; (4) AL ＞ 24 mm. The Vision Transformer (ViT) deep learning model was used to predict the length interval of AL, and pre-processing methods such as Gaussian filter combined with transfer learning were used to improve the learning ability of the model. In order to more fully verify the performance of the model, the K-fold cross-validation method combined with ROC curve, area under ROC curve (AUC), precision and recall rate were used to evaluate the comprehensive performance of model.

Results The AUC values of the model ranged from 0.750 to 0.883 under different AL, and the AUC of ＞ 24 mm was the largest (0.883). The overall accuracy of the model in predicting the AL interval was about 0.633, the accuracy range of the model was 0.360–0.802 under different AL intervals, and the recall rate was 0.294–0.802. The accuracy and recall rate of the model with the AL interval ＞24 mm were both 0.802, which showed high accuracy and high reliability. The heat maps showed that the model focused more on posterior polar atrophy in the fundus region with lesions.

Conclusions It is basically feasible to predict the AL interval of children's eyes through deep learning model based on UWF images.