Hybrid deep learning convolutional neural networks and optimal nonlinear support vector machine to detect presence of hemorrhage in retina

Abstract Diabetic retinopathy is a disorder that occurs in retina and it is caused by diabetes mellitus. Millions of people with diabetic retinopathy are expected to experience a loss of vision across the globe. Therefore, accurate automated-diagnosis systems are highly needed to help physicians in clinical milieu. Though many factors are effective in the diagnosis of diabetic retinopathy, presence of hemorrhage in retina remains one of the most significant factors. We present a three-stage hybrid system for classification of normal and abnormal digital retina images with hemorrhage. First, deep learning convolutional neural networks (CNN) is used for automatic features extraction. Second, the Student t-test is applied to the high dimensional features set extracted by CNN to select the best ten features. Third, the selected CNN-based features are fed to a nonlinear support vector machine (SVM) tuned by Bayes optimization to perform classification task. Three additional popular classifiers are also trained with features extracted by CNN and their performances are compared to that of the optimal nonlinear SVM including linear discriminant analysis (LDA), naive Bayes (NB), and k nearest neighbor (kNN). Each automated-diagnosis system is validated on a database composed of healthy and unhealthy digital retina images affected with various grades of hemorrhage. Experimental results from ten-fold cross-validation methodology show that CNN-SVM outperforms all other three reference systems; namely, CNN-LDA, CNN-NB, and CNN-kNN. Indeed, CNN-SVM system achieved 99.11%±0.0101 accuracy, 99.14%±0.0143 sensitivity, 99.08%±0.0083 specificity, and 0.97.31%± 0.0381 area under curve (AUC) of the receiver operating characteristic. The proposed system is fast and accurate.