Smart detection and diagnosis of diabetic retinopathy using bat based feature selection algorithm and deep forest technique

Abstract

Diabetic retinopathy is a retinal eye disease due to presence of diabetes and diabetes can be described as metabolic disorder in which glucose level is higher in human body. Diabetic retinopathy can also be responsible for vision loss in people with diabetes and final stage of diabetic retinopathy is complete blindness. The earlier detection of DR can be crucial for preventing the visual disturbances. Hence, the aim of this work is to develop an efficient model for early detection and diagnosis of diabetic retinopathy using fundus images. The working of the proposed model described through four phases- preprocessing, feature extraction, feature selection and classification. The preprocessing phase corresponds to removal of noise and enhance the contrast of images. The noise are eliminated through noise channel and median filter. Further, CLAHE algorithm is chosen for improving the contrast of images. The lesion region is identified through k-mean based segmentation algorithm. In feature extraction phase, multi-grained scanning method is applied for extracting the high dimensional features from enhanced images. These high dimensional features are given to bat algorithm for determining relevant and optimized features for detection of diabetic retinopathy, called bat based feature selection algorithm. Finally, the relevant features are considered for the classification and diagnosis of diabetic retinopathy using DeepForest cascade technique, called BA-DeepForest and task of this technique provides the binary classification of diabetic retinopathy. The efficacy of proposed BA-DeepForest model is evaluated using two diabetic retinopathy datasets. The simulation results are compared with several existing models like KNN, ANN, SVM, VGG16, VGG19, InceptionV3 and Deep Forest. The results showed that proposed BA-DeepForest model achieves higher accuracy rates (92.94 & 94.61), F1-score rates (95.31 & 96.49), sensitivity rates (95.67 & 97.39), and specificity rates (94.95 & 98.04) using 10-cross fold method.