A few-shot learning-based eye diseases screening method.

OBJECTIVE: This study aims to construct a brand-new ophthalmic disease screening task and establish a practically valuable ophthalmic disease screening model in the case of insufficient data. MATERIALS AND METHODS: The main methods are as follows: firstly, we mixed data from different sources (these data may come from different cameras, including different fundus diseases) to get a new dataset. Based on this dataset, we conducted subsequent experiments on fundus multi-disease screening. However, in the past public datasets, each dataset often only corresponded to the screening diagnosis of one disease. Secondly, we proposed a method to simulate the characteristics of different fundus cameras by using a method based on style transfer, and to augment the training data, so that the model could learn the features of ophthalmic diseases in a more comprehensive way. Finally, a robust disease screening model based on few-shot learning was constructed on the combined dataset, and compared with benchmark algorithms. RESULTS: We focused on the study of eye disease screening methods based on the metric-based few-shot learning model, data augmentation methods, and focus on key technologies such as data augmentation based on style transfer. Experiments have shown that our method can significantly improve the generalization ability of the disease screening model. CONCLUSIONS: By introducing few-shot learning theory and data augmentation based on style transfer into ophthalmic disease screening, the generalization ability of the model is greatly improved, and it has certain practical value. [ABSTRACT FROM AUTHOR]