A method and device for identifying diabetes mellitus disease with fusion of binocular features

Abstract

The invention discloses a method and device for identifying diabetic retinopathy by integrating binocular features, comprising: acquiring a plurality of medical fundus picture groups and performing preprocessing respectively; Extraction, dimensionally superimpose the extracted left-eye feature map and right-eye feature map to achieve feature fusion, perform global maximum pooling on the fused feature map, use the dropout layer to reduce the overfitting of the model, and then feature Carry out full connection to the feature vector with a length of 2K, and then separate it into K-dimensional left-eye features and right-eye features again, divide the left-eye features and right-eye features into two subcategories, and train them with labels respectively, and get The left-eye identification sub-model and the right-eye identification sub-model are used to identify the degree of diabetic retinopathy in the left eye and the right eye. The invention strengthens the network's attention to the common features of the left and right eyes through feature fusion, thereby simultaneously improving the category recognition accuracy of the two pictures.

Description

technical field [0001] The invention relates to the field of image processing, and more specifically, to a method and device for identifying diabetic retinopathy by fusing binocular features. Background technique [0002] At present, the recognition of diabetic retinopathy images is mainly divided into feature recognition methods for specific lesions and global classification methods for images. The feature recognition method of specific lesions is mainly to separate and identify specific lesions such as hemorrhage and flocculent spots through mathematical morphology and image processing technology. The global classification of images is to classify a large number of marked images through convolutional neural networks. The data is trained, and then the image to be recognized is automatically classified through the trained model. [0003] For the identification method of specific lesions, such as Haloi et al., first perform anisotropic diffusion filtering method on the image...

AI Technical Summary

Problems solved by technology

[0005] The technical problem to be solved by the present invention is that the problem faced by the current mainstream network structure in the prior art is that its input data often only has one image, and there are certain limitations for the data of diabetes mellitus disease, which has both left and right eye images. , resulting in the technical defect that the recognition accuracy is not high enough, providing a method and device for recognizing diabetic reticulopathy that integrates binocular features

Images