Robust nonlinear multimodal classification of Alzheimer's disease based on GMM

Abstract

Accurate \u00a0diagnosis \u00a0of \u00a0Alzheimer's \u00a0disease \u00a0(AD) \u00a0and \u00a0its \u00a0prodromal \u00a0stage \u00a0mild \u00a0cognitive impairment \u00a0(MCI) \u00a0is \u00a0very \u00a0important \u00a0for \u00a0patients \u00a0and \u00a0clinicians. \u00a0There \u00a0are \u00a0many \u00a0useful \u00a0medical \u00a0data \u00a0have been discovered to be remarkable for diagnosis i.e., structural MR imaging (MRI), functional imaging (e.g., FDG-PET and FIB-PET). Multimodal classification model is needed to combine these biomarkers to improve the \u00a0diagnose \u00a0performance. \u00a0Some \u00a0methods \u00a0have \u00a0been \u00a0proposed \u00a0such \u00a0as \u00a0linear \u00a0mixed \u00a0kernel, \u00a0combined embedding and nonlinear graph fusion. These methods have efficiently employed the multimodal data, but they ignore the influence of noise and outliers. Noise is easily generated in image analysis and measurement. To enhance robustness, mixture distributions were applied in nonlinear regression models. Gaussian mixture model \u00a0is \u00a0successfully \u00a0applied \u00a0in \u00a0many \u00a0domains. \u00a0In \u00a0this \u00a0paper, \u00a0we \u00a0generalize \u00a0nonlinear \u00a0multimodal classification model based on GMM. The performance on real dataset: 22 AD, 23 MCI and 25 NC (health) is comparable to other methods.