Automatic grading method for diabetic retinopathy image based on label coding

Abstract

The invention discloses an automatic grading method for a diabetic retinopathy image based on label coding, and the label coding technology not only can be used for grading the diabetic retinopathy image, but also can be applied to other orderly classification problems. According to the method, firstly, soft coding is carried out on five lesion grade labels of the diabetic retinopathy images to control model prediction probability distribution so as to solve the class dependence problem among the labels; a diabetic retinopathy data set is built and model parameters and offset parameters in theparameter soft labels are obtained through model training; soft codes of five lesion levels are calculated by utilizing the offset parameters, wherein the soft codes can be used for transfer learning; and finally, accurately the diabetic retinopathy images are graded accurately through the model obtained by training. The method can effectively eliminate the dependence between labels in the diabetic retinopathy grading problem, the label prediction probability distribution of the model can be flexibly controlled through the soft coding method, and the accuracy of diabetic retinopathy grading is improved; and meanwhile, hard and soft label mapping can be established for transfer learning.

Description

Technical field[0001]The present invention relates to the application field of image processing and machine vision, especially the severity assessment and intelligent grading of diabetic-induced retinopathy. The label coding technology in the present invention is also applicable to other ordered classification problems in deep learning.Background technique[0002]In the severity classification of diabetes-induced retinopathy, according to the International Clinical Diabetic Retinopathy (DR) Severity Scale, as shown in Table 1, the severity of the disease can be divided into five categories: normal, mild, moderate, severe, and value-added, such asfigure 1 Shown. Part of the classification of the severity of the lesion is an ordered classification problem, and there is a dependency relationship between each category. For a certain label, different categories play different roles in predicting probability distribution. For example, different predictions for value-added (DR4) fundus photo...

AI Technical Summary

Problems solved by technology

The disadvantage of this algorithm is that it can only convert hard labels into fixed soft label distributions, and does not have generalization ability.

Images