Integral correlation based optimal SD-OCT slice selection and pathology classification using handcraft and deep features

Abstract

In this paper a novel classification system IC_HBO_DF (Integral Correlation Haar-like Biorthogonal Deep Features) for automatically classifying the retinal diseases of Age-related Macular Degeneration (AMD) and Diabetic Macular Edema (DME). The Spectral Domain Optical Coherence Tomography (SD-OCT) images are taken from the DUKE dataset that contains 15 AMD, 15 DME and 15 normal. Integral correlation function is applied to generate the optimized slices. Haar-like feature and biorthogonal approximation coefficient are extracted from the optimized slices. The convolutional neural network based pretrained GoogLeNet and VGGNet16 are used to extract the deep features from the Haar-like feature and biorthogonal approximation coefficient. Machine learning techniques Random forest, LB_Boost, DecisionTree, Adaptive_Boost are used to classify the retinal pathologies of AMD and DME. Experimental results demonstrate the effectiveness of the proposed VGGNet16 based Random forest classification algorithm that has achieved higher degree of accuracy.