Image Enhancement and 27 Pretrained Convolutional Neural Network Models for Diabetic Retinopathy Grading

Abstract

Diabetic retinopathy (DR) affects the retina's blood vessels and causes vision loss. Fundus images are used to diagnose DR, which is a lengthy process because experienced clinicians must accurately diagnose the disease and identify microlesions early to prevent blindness. Computer vision can be used for retinal image classification. The APTOS dataset contains 5990 normal, moderate, mild, proliferate, and severe retinal images. In this study, we proposed a convolutional neural network (CNN) ensemble for DR fundus grading. Each image channel was enhanced by contrast-limited adaptive histogram equalization (CLAHE) and gamma correction and then fed to 27 pretrained CNN models for one-time training to examine the DR grading. The results showed that MobileNet's green channel with the CLAHE technique is sufficiently fast and accurate for disease classification. The grading retinal images had an accuracy of 96.95%, a precision of 96.17%, a sensitivity of 97.80%, an F1 score of 96.98%, and a specificity of 97.75%. In addition, the proposed method improves the speed and robustness of retinal DR grading.