High myopia fundus lesion risk prediction method

Abstract

The invention discloses a high myopia fundus lesion risk prediction method based on a convolutional neural network and knowledge distillation. The risk prediction method comprises the steps that fundus images are acquired, and data of a training set and a test set are randomly divided; preprocessing such as random overturning, cutting, color jittering and normalization is performed on the trainingset, and only normalization preprocessing is performed on the test set; a classification network model is trained by using a knowledge distillation method, and training data are correspondingly sentto a pre-trained teacher network and a to-be-trained student network; the soft label value and the real label value output by the teacher network are taken as supervision information, and KL Loss andFocal Loss are calculated according to the supervision information and predicted values output by the student network; the two different Loss values are weighted and summed to serve as a final loss function for parameter updating of the student network; and the trained student network can perform three-classification prediction of normal-low-risk high myopia fundus lesion-high-risk high myopia fundus lesion on the fundus image test set.

Description

technical field [0001] The invention relates to a method for predicting the risk of fundus lesions in high myopia based on convolutional neural network and knowledge distillation, which belongs to the field of medical image (fundus image) processing. Background technique [0002] my country is a country with a high incidence of myopia, and the number of patients with high myopia (myopia degree above 600 degrees) is increasing year by year, and it is showing a younger trend. High myopia may cause a variety of serious complications, even blindness, and the vision damage it brings is permanent and irreversible. It is currently recognized that high myopia is a blinding eye disease. According to the "Expert Consensus on the Prevention and Control of High Myopia" formulated by the Ophthalmology Group of the Chinese Medical Association Ophthalmology Branch in 2017, high myopia can be divided into low-risk simple high myopia and high-risk pathological myopia. Although low-risk high...

AI Technical Summary

Problems solved by technology

However, artificial high myopia risk prediction is time-consuming and laborious, and the accuracy is not high, so it is difficult to promote and implement

In areas where medical resources are relatively scarce, it is more difficult for patients with high myopia to receive regular follow-up monitoring. The shortage of ophthalmologists and medical equipment may cause patients with high myopia to miss the window period for optimal prevention and treatment

Images