Optimized Deep Learning based Approach for Enhanced frame work of Automated Diagnosis of Diabetic Retinopathy

Abstract

Listen

Translate article

Diabetic Retinopathy is a major threat to cause vision loss in people suffering from Diabetes Mellitus. Many machine learning algorithms were proposed to detect Diabetic Retinopathy (DR) at an early stage, and with proper treatment vision loss may be reduced. This paper proposes a novel method to detect DR through severity scale by observing the abnormalities through ensemble methods. Deep learning based models are gaining focus to construct automated tools for medical image analysis. This paper uses Alex Net based DNN (Deep Neural Network) which functions on the basis of Convolution Neural Network (CNN) and is applied to have an optimal solution for automated DR detection with Random Forest Classifier (RFC). Recursively Separated and Weighted Histogram Equalisation (RSHWE) is used to preserve brightness, ensemble of segmentation algorithms to the identify Region of Interest (ROI). Feature map constructed using Gaussian and Gabor filter coefficients and Grey Level Co occurrence Matrix (GLCM) features and these features are applied to Random Forest Classifier (RFC) to classify the diseased images. The performance of RFC is also compared with and without Gradient features with Enhanced RFC (E-RFC). The accuracy of various classifiers is compared with our proposed method. In this paper, the considered performance metrics are accuracy, sensitivity, specificity. This method experimented on publicly available fundus image data sets for DR and shows good results with an accuracy (94.8%), specificity (93%), sensitivity (96%).