SUMMARY
Diabetic retinopathy (DR) is one of the most common vascular problems. It causes due to the abnormal growth of blood vessels that may lead to blindness if it is not treated properly in early stages. Usually, the early signs cannot be seen by manual intervention until it reaches an advanced stage as well as differentiating between Non-proliferative and Proliferative is quite difficult. To handle this, Computer Aided Diagnosis (CAD) considered as an impressive way to decrease the rate of sufferers through recognizing the level of the damages according to Non-proliferative and Proliferative.
In this work, four different methods were proposed for screening and classifying of 150 downloaded fundus images from Digital Retinal Images for Vessel Extraction (DRIVE) project. The images are equally distributed into Normal, NON-proliferative (NONPDR) and Proliferative (PDR) stages. In the first method, a second order Statistical Based Texture Features (SBTF) were applied to extract the texture of the regions. The main features are: Gray Level Cooccurrence Matrix (GLCM), Gray Level Run Length Matrix (GLRLM), and difference statistics features. For the sake of improving the performance of first method three other methods were proposed: In the second method, Discrete Wavelet-Based Statistical Features (DWBSF) was applied before extraction the features. Here, the images transformed by using 2nd level decomposition of two-dimensional discrete wavelet transform technique. In the third method, Gaussian Matched Filter (GMF) was performed in order to extract the Vessel Segmentation area Based Statistical Features (VSBSF) from the prepared images after pre-processing step. In the last method, Neutrosophic Based Statistical Features (NBSF) were extracted from the pre-process images. At first, an image is transformed to Neutrosophic Set (NS) domain that has True, Indeterminacy, and False subsets and then, statistical features were extracted.
From all proposed methods, the extracted features were tested by using a features selection scheme named One-Way Analysis of Variance (ANOVA1) with 0.001 level of significance. The successive features were fed to Naïve Bayes (NB), Support Vector Machine (SVM), K-Nearest Neighbour (KNN) classifiers. The final performance results were improved by conducted neural network with Particle Swarm Optimization (PSO) and Genetic Algorithm (GA).
The results from VSBSF and NBSF showed that the system robustness in classification DR stages within the average: accuracy 96.7%, sensitivity 98% and specificity 96%, for NBSF and accuracy 89.3%, sensitivity 87,2%, and specificity 91.3% for VSBSF. The results of these two methods were compared with results of some previous works. A comparison illustrates these two proposed methods can be used in clinical observation to help the doctors to improve their decision about the patient’s case and reduce the number of blindness.