Automated detection of age-related macular degeneration using empirical mode decomposition

Fundus images are converted to 1D signals using Radon transform.Empirical Mode Decomposition (EMD) is applied on the 1D signal.Nonlinear features are extracted from IMFs derived from EMD.Various ranking methods are used to identify optimal features.Classification accuracy of 100% is obtained for STARE database. Age-related Macular Degeneration (AMD) is the posterior segment eye disease affecting elderly people and may lead to loss of vision. AMD is diagnosed using clinical features like drusen, Geographic Atrophy (GA) and Choroidal NeoVascularization (CNV) present in the fundus image. It is mainly classified into dry and wet type. Dry AMD is most common among elderly people. At present there is no treatment available for dry AMD. Early diagnosis and treatment to the affected eye may reduce the progression of disease. Manual screening of fundus images is time consuming and subjective. Hence in this study we are proposing an Empirical Mode Decomposition (EMD)-based nonlinear feature extraction to characterize and classify normal and AMD fundus images. EMD is performed on 1D Radon Transform (RT) projections to generate different Intrinsic Mode Functions (IMF). Various nonlinear features are extracted from the IMFs. The dimensionality of the extracted features are reduced using Locality Sensitive Discriminant Analysis (LSDA). Then the reduced LSDA features are ranked using minimum Redundancy Maximum Relevance (mRMR), Kullback-Leibler Divergence (KLD) and Chernoff Bound and Bhattacharyya Distance (CBBD) techniques. Ranked LSDA components are sequentially fed to Support Vector Machine (SVM) classifier to discriminate normal and AMD classes. The performance of the current study is experimented using private and two public datasets namely Automated Retinal Image Analysis (ARIA) and STructured Analysis of the Retina (STARE). The 10-fold cross validation approach is used to evaluate the performance of the classifiers and obtained highest average classification accuracy of 100%, sensitivity of 100% and specificity of 100% for STARE dataset using only two ranked LSDA components. Our results reveal that the proposed system can be used as a decision support tool for clinicians for mass AMD screening.