Combining MF-DFA and LSSVM for retina images classification

Diabetic retinopathy is the main cause of blindness in adults. Early diagnosis of diabetic retinopathy is essential for avoiding deterioration of illness and vision loss. The use of computer technology to identify diabetic retinopathy images provides significant means to reduce the risk of deterioration. In this paper, we propose a new approach for retina images detection and classification by using a hybrid system which is constructed by two-dimensional multifractal detrended fluctuation analysis (2D MF-DFA) and least square support vector machines (LSSVM). In the proposed method, we applied 2D MF-DFA to compute the local generalized Hurst exponents which are the multifractal features of the diabetic retinopathy image, and these values are recorded as LH_q. Then, the Hurst exponents are taken as the training input vector for the training in LSSVM. Finally, we classified a specific retina image as healthy or lesion image. We present experimental verification to investigate the efficiency and robustness of the proposed system. The results show that the proposed system yields a classification accuracy with 99.01% ± 0.0074, sensitivity with 99.03% ± 0.0051, and specificity with 97.73% ± 0.0075. When the performance was compared with state-of-the-arts, the solution indicated that the MF-DFA-LSSVM system outperforms most of others in terms of all the classification sensitivity, accuracy, and specificity. The proposed method will be useful for clinical medicine.