COPD DETECTION USING THREE-DIMENSIONAL GAUSSIAN MARKOV RANDOM FIELDS BASED ON BINARY FEATURES

This paper proposes new descriptors based on three-dimensional Gaussian Markov random fields (3D-GMRF) for volumetric texture classification. The estimated parameters of 3D-GMRF are decomposed into sign and magnitude components and then are encoded into a single binary code to describe the local texture. Our experiments on a synthetic dataset of volumetric texture show that this approach leads to significant reduction in descriptor size, while preserving the discriminative power of 3D-GMRF features. The descriptors proposed here demonstrate strong performance in distinguishing between healthy and chronic obstructive pulmonary disease (COPD) subjects, using a medical dataset. These descriptors are successfully employed to measure the differences between various groups from the medical dataset, in order to determine which group is at risk of COPD.