Performance analysis of descriptive statistical features in retinal vessel segmentation via fuzzy logic, ANN, SVM, and classifier fusion
文摘
Diabetic retinopathy is the most common diabetic eye disease and a leading cause of blindness in the world. Diagnosis of diabetic retinopathy at an early stage can be done through the segmentation of blood vessels of the retina. In this work, the performance of descriptive statistical features in retinal vessel segmentation is evaluated by using fuzzy logic, an artificial neural network classifier (ANN), a support vector machine (SVM), and classifier fusion. Newly constructed eight features are formed by statistical moments. Mean and median measurements of image pixels’ intensity values in four directions, horizontal, vertical, up-diagonal, and down-diagonal, are calculated. Features, F1, F2, F3, and F4 are calculated as the mean values and F5, F6, F7, and F8 are calculated as the median values of a processed pixel in each direction. A fuzzy rule-based classifier, an ANN, a SVM, and a classifier fusion are designed. The publicly available DRIVE and STARE databases are used for evaluation. The fuzzy classifier achieved 93.82% of an overall accuracy, 72.28% of sensitivity, and 97.04% of specificity. For the ANN classifier, 94.2% of overall accuracy, 67.7% of sensitivity, and 98.1% of specificity are achieved on the DRIVE database. For the STARE database, the fuzzy classifier achieved 92.4% of overall accuracy, 75% of sensitivity, and 94.3% of specificity. The ANN classifier achieved the overall accuracy, sensitivity, and specificity as 94.2%, 56.9%, and 98.4%, respectively. Although the overall accuracy of the SVM is calculated lower than the fuzzy and the ANN classifiers, it achieved higher sensitivity rates. Designed classifier fusion achieved the best performance among all by using the proposed statistical features. Its overall accuracy, sensitivity, and specificity are calculated as 95.10%, 74.09%, 98.35% for the DRIVE and 95.53%, 70.14%, 98.46 for the STARE database, respectively. The experimental results validate that the descriptive statistical features can be employed in retinal vessel segmentation and can be used in rule-based and supervised classifiers.
