Detection of pathologies in retina digital images an empirical mode decomposition approach

Abstract

Accurate automatic detection of pathologies in retina digital images offers a promising approach in clinicalapplications. This thesis employs the discrete wavelet transform (DWT) and empirical mode decomposition (EMD) to extract six statistical textural features from retina digital images. The statistical features are the mean, standard deviation, smoothness, third moment, uniformity, and entropy. The purpose is to classify normal and abnormal images. Five different pathologies are considered. They are Artery sheath (Coat’s disease), blot hemorrhage, retinal degeneration (circinates), age-related macular degeneration (drusens), and diabetic retinopathy (microaneurysms and exudates). Four classifiers are employed; including support vector machines (SVM), quadratic discriminant analysis (QDA), k-nearest neighbor algorithm (k-NN), and probabilistic neural networks (PNN). For each experiment, ten random folds are generated to perform cross-validation tests. In order to assess the performance of the classifiers, the average and standard deviation of the correct recognition rate, sensitivity and specificity are computed for each simulation. The experimental results highlight two main conclusions. First, they show the outstanding performance of EMD over DWT with all classifiers. Second, they demonstrate the superiority of the SVM classifier over QDA, k-NN, and PNN. Finally, principal component analysis (PCA) was employed to reduce the number of features in hope to improve the accuracy of classifiers. We find that there is no general and significant improvement of the performance, however. In sum, the EMD-SVM system provides a promising approach for the detection of pathologies in digital retina.

Date	15 Dec 2011
Original language	American English
Awarding Institution	École de technologie supérieure
Supervisor	Christian Gargour (Supervisor) & Gheorghe Marcel Gabrea (Co-supervisor)