Neuronet modeling in the implementation of the system for determining retinal damage of diabetic origin

Abstract

In order to determine the stage of retinal damage of diabetic origin, machine learning mechanisms are applied. The use of the DenseNet convolutional neural network for high-quality image recognition and segmentation is substantiated. DenseNet-121, DenseNet-169 and DenseNet-201 networks have been modified by adding additional layers. Software mechanisms for image processing using Gaussian blurring, removal of black frames, and minimization of the influence of image position changes on recognition quality have been developed. The model was built and trained. High rates of recognition accuracy were obtained. For the DenseNet-201 network, an indicator of 97.9% was obtained, which exceeds the characteristics of the DenseNet-121 and DenseNet-169 networks. Figs.: 2. Tabl.: 1. Refs.: 13 titles.