Fundus Image Recognition Based on Deep Learning: Methodology and Experimental Analysis
Main article
Abstract
Fundus is an important research object in ophthalmology. Ophthalmologists often diagnose human eye diseases through fundus images. To improve the diagnosis rate of eye diseases by ophthalmologists, the feasibility of using image processing and artificial intelligence technology to intelligently identify fundus images is studied. The fundus image data is effectively processed using a variety of image technologies included in OpenCV and MATLAB. Convolutional neural networks with different structures and parameters are established using TensorFlow to extract features and train the processed fundus images. The recognition rate is improved by changing the hierarchy and adding the center loss function to optimize the network. The fundus images are input into the network to achieve rapid classification of fundus diseases and display the corresponding recognition probability. 500 250*166-pixel fundus images of eight categories including diabetes and glaucoma are selected for classification training and test experiments. The results show that the recognition rate of the network model reaches 44.81% when the number of iterations is 20,000. This method realizes the recognition of fundus images and can help ophthalmologists assist in the judgment of eye diseases.
