Intelligent Diabetic Retinopathy Detection using Deep Learning

Abstract

Diabetic retinopathy (DR) is the most common illness related to diabetes caused by the increasing of glucose in human blood and has been dramatically increased in the last decade. Practically, DR is examined by conducting manual analysis on retina images resulted from fundus camera modality in which can lead to some problems such as time-consuming, need more thoroughness and properly skill and experience. Due to the insufficient number of ophthalmologists, especially in rural areas, an alternative solution in supporting diagnosis properly is needed. Regarding to those issues, some research communities have proposed intelligent system for detecting DR. Despite some previous intelligent DR detection have been developed, there still remained problem that quality of image was extremely affect the performance. Hence, in this study we proposed an intelligent DR detection completed with image enhancement process for maintaining the model performance. Our proposed solution was performed in 200 retina images consisting of two classes (normal and abnormal or DR). Our proposed solution successfully increased the performance with the highest accuracy, sensitivity, specificity, positive predictive value, and negative predictive value of 0.92, 0.95, 0.81, 0.95, 0.81, respectively. This result has increased by around of 40% in most of evaluation metrics of the model's performance without an image enhancement process. It indicates that conducting image enhancement process before training the model was important to increase the model performance and to prevent the miss-detection.