Development of diabetic retinopathy early detection and its implementation in Android application

Abstract

Diabetic retinopathy (DR) is a diabetes complication causing blindness in which symptoms are not perceived in earlier stage or non-proliferative diabetic retinopathy (NPDR). It is difficult for manual diagnosis methods to keep pace with the growing number of DR. In this study, an algorithm to detect NPDR was developed and implemented in the Android application. In contrary to feature engineering, this study explored a different classification approach by having used a deep neural networks and transfer learning methods on fundus images to train the classifier models. Model development utilized Messidor (4 class) dataset and Messidor-2 (2 class) dataset, image pre-processing, Inception V3 network and MobileNetV1 network, the configuration of test set-train set split, optimizer, and learning rate. Test accuracy of 86% was acquired with InceptionV3 and Messidor-2 which then implemented in Android application. Its yielded accuracy, sensitivity, and specificity are 88%, 80%, and 76% respectively.Diabetic retinopathy (DR) is a diabetes complication causing blindness in which symptoms are not perceived in earlier stage or non-proliferative diabetic retinopathy (NPDR). It is difficult for manual diagnosis methods to keep pace with the growing number of DR. In this study, an algorithm to detect NPDR was developed and implemented in the Android application. In contrary to feature engineering, this study explored a different classification approach by having used a deep neural networks and transfer learning methods on fundus images to train the classifier models. Model development utilized Messidor (4 class) dataset and Messidor-2 (2 class) dataset, image pre-processing, Inception V3 network and MobileNetV1 network, the configuration of test set-train set split, optimizer, and learning rate. Test accuracy of 86% was acquired with InceptionV3 and Messidor-2 which then implemented in Android application. Its yielded accuracy, sensitivity, and specificity are 88%, 80%, and 76% respectively.