Co-training based semi-supervised classification of Alzheimer´s disease

The Alzheimer´s disease (AD) is the most prevalent neurodegenerative brain disease worldwide. The neuroimaging based computer-aided diagnosis (CAD) of AD has attracted much attention. Therefore, machine learning plays an important role in it. However, it is usually a time-costing task to label a large-scale data in clinical practice, which results in the neuroimaging dataset for AD is a small sample size. The semi-supervised learning (SSL), which has the ability to employ the easily acquired unlabeled data to assist insufficient labeled data for improving learning performance, have attracted considerable interest for AD classification. The co-training classification is a representative paradigm of disagreement-based SSL method. It trains alternately to maximize the mutual agreement on two preferably independent views (feature subsets) of the unlabeled data. Since the magnetic resonance imaging (MRI) and positron emission tomography (PET) are the commonly used neuroimaging techniques for AD diagnosis, co-training method has the potential to perform semi-supervised classification of AD with both MRI and PET data. In this paper, we propose the co-training based SSL with MRI and PET images for classification of AD with the mild cognitive impairment (MCI) data as the unlabeled samples. We compare the co-training method with other supervised and semi-supervised classification method on the ADNI dataset. The experimental results indicate that co-training method with both MRI and PET data achieves best performance, which suggest that it has the potential to be applied to neuroimaging based CAD for AD.