Effective Diagnosis of Alzheimer's Disease by Means of Distance Metric Learning

In this paper we present a novel classification method of SPECT images for the early diagnosis of the Alzheimer's disease (AD). The proposed method is based on distance metric learning classification with the Large Margin Nearest Neighbour algorithm (LMNN) aiming to separate examples from different classes (Normal and AD) by a large margin. In particular, we show how to learn a Mahalanobis distance for k-nearest neighbors (KNN) classification. It is also introduced the concept of energy-based model which outperforms both Mahalanobis and Euclidean distances. The system combines firstly Normalized Minimum Square Error (NMSE) and t-test selection with secondly Kernel Principal Components Analysis (KPCA) to find the main features. Applying KPCA trick in the feature extraction, LMNN turns into Kernel-LMNN (KLMNN) with better results than the first. KLMNN reachs results of accuracy=96.91%, sensitivity=100%,specificity=95.35% outperforming other recently reported methods such as Principal Component Analysis(PCA) in combination with Linear Discriminant Analysis (LDA) evaluated with Support Vector Machines (SVM) or linear SVM.