Comparison of Image Pre-processing for Classifying Diabetic Retinopathy Using Convolutional Neural Networks

Diabetes mellitus (DM) is a global health problem that results in different conditions, and one of the most problematic is diabetic retinopathy (DR), as it may have no symptoms in its early stages and can leave the patient completely blind. Some authors have created different convolutional neural network (CNN) models for the detection and classification of DR and thus help experts when deciding the best treatment for the patient. To create a CNN model, it is desirable to pre-process the dataset to improve image classification accuracy. For this reason, this work aims to compare the mean accuracy of two CNN models: the first one using three convolution layers, while the second one uses ten layers. For this work, to test the proposed models, the APTOS 2019 database is used with four different pre-processing types. The importance of applying pre-processing is reflected in the improvement of the precision results obtained by the CNNs models. Because of this, it is preferable to apply the best possible pre-processing to the database. For the comparisons, the work was carried out with two case studies: binary and multiclass (five stages of DR). The obtained results were that, for the binary case, the highest mean accuracy obtained was by the second pre-processing (using the CNN model of depth 3) with 0.94. In the case of multiclass problem, the best mean accuracy was also obtained by the second pre-processing (using the CNN model of depth 3) with 0.74.