Abstract
Incorporating with machine learning technology, neuroimaging markers which extracted from structural Magnetic Resonance Images (sMRI), can help distinguish Alzheimer’s Disease (AD) patients from Healthy Controls (HC). In the present study, we aim to investigate differences in atrophic regions between HC and AD and apply machine learning methods to classify these two groups. T1-weighted sMRI scans of 158 patients with AD and 145 age-matched HC were acquired from the ADNI database. Five kinds of parameters (i.e. cortical thickness, surface area, gray matter volume, curvature and sulcal depth) were obtained through the preprocessing steps. The recursive feature elimination (RFE) method for support vector machine (SVM) and leave-one-out cross validation (LOOCV) were applied to determine the optimal feature dimensions. Each kind of parameter was trained by SVM algorithm to acquire a classifier, which was used to classify HC and AD ultimately. Moreover, the ROC curves were depicted for testing the classifiers’ performance and the SVM classifiers of two-dimensional spaces took the top two important features as classification features for separating HC and AD to the maximum extent. The results showed that the decreased cortical thickness and gray matter volume dramatically exhibited the trend of atrophy. The key differences between AD and HC existed in the cortical thickness and gray matter volume of the entorhinal cortex and medial orbitofrontal cortex. In terms of classification results, an optimal accuracy of 90.76% was obtained via multi-parameter combination (i.e. cortical thickness, gray matter volume and surface area). Meanwhile, the receiver operating characteristic (ROC) curves and area under the curve (AUC) were also verified multi-parameter combination could reach a better classification performance (AUC=0.94) after the SVM-RFE method. The results could be well prove that multi-parameter combination could provide more useful classified features from multivariate anatomical structure than single parameter. In addition, as cortical thickness and multi-parameter combination contained more important classified information with fewer feature dimensions after feature selection, it could be optimum to separate HC from AD to take the top two important features of them to construct SVM classifiers in two-dimensional space. The proposed work is a promising approach suggesting an important role for machine-learning based diagnostic image analysis for clinical practice.
Brain development during adolescence: A mixed-longitudinal investigation of cortical thickness, surface area, and volume
