A Deep Learning Strategy Using the CNN Model DRDnet22 for the Early Detection of Diabetic Retinopathy From Retinal Images

Abstract

Diabetic Retinopathy, or DR, is a major problem with the eyes of diabetes patients. If the DR is found early, many people can avoid going blind and other complications. Several systems based on artificial intelligence have been proposed, and they are better at finding the DR than human analysis in terms of time efficiency and human dependencies. Manual screening using retinal fundus images, such as visual acuity testing, pupil dilation, and optical consistency tomography, requires highly skilled clinicians to find and evaluate the importance of many small details. This is a difficult, time-consuming, and error-prone task. Because of this, a computer-aided, automated process is a must. In this thesis, the APTOS 2019 dataset is used for training and testing. Which is made up of 3662 named pieces of data. DRDnet22 is a model that uses CNN to find and classify DR into 5 phases based on the severity level. Since each ConvNet gets different features, combining them with 1-D pooling and cross-pooling gives a better representation than just using the features from one ConvNet. For comparison, the traditional pre-trained model was also trained. Evaluation of performance indicates that the proposed model was more accurate with 81.6% than the traditional pre-trained models.