Fundus image lesion degree identification and visualization system based on convolutional neural network

Abstract

The invention provides a fundus image lesion degree identification and visualization system based on a convolutional neural network. The system is characterized by comprising a data inter-class exponential form sampling module in a training stage, a uniform inter-sample sampling module in the same class in the training stage, an image preprocessing module in the training stage, an image preprocessing module in a test stage and a fundus image prediction module in the test stage. The sampling method provided by the invention not only is simple to implement, but also is much faster in sampling speed. According to the neural network structure provided by the invention, a detection result is better than that of an existing neural network structure, and a discriminative region can be visualizedunder the condition that only an image-level label is used.

Description

technical field [0001] The invention relates to a recognition and visualization system of fundus image lesion degree based on a convolutional neural network, which belongs to machine learning, image processing, and computer vision technologies. Background technique [0002] With the continuous development of image processing, computer vision, deep learning and other technologies, fundus images have also been extensively studied. The four aspects of data preprocessing, category imbalance mechanism, uniform sampling method of samples within a category, and neural network structure design are the key points of the convolutional neural network fundus image lesion detection method. Data overfitting strategies can improve the generalization of convolutional neural network models. The category imbalance mechanism allows the model to keep focusing on a small number of categories instead of only learning the features of a large number of categories. The uniform sampling method of s...

AI Technical Summary

Problems solved by technology

Such methods are applicable to large-scale datasets, but cannot localize discriminative regions in fundus images

Images