Diagnosis of Alzheimer’s disease by structural MRI-Validation of efficiency of AI-derived Alzheimer’s disease score

Abstract

Voxel-based morphometry (VBM) uses structural MRI data to investigate brain region volumes in a voxel-wise manner, not unlike computing z-scores in SPECT using eZIS or iSSP. Recently, we added artificial intelligence (AI) to our software “BAAD” (Brain Anatomical Analysis using Diffeomorphic deformation) that was originally developed to support diagnosis of Alzheimer’s disease (AD). The AI combines support vector machine (SVM) with a radial basis function (RBF) kernel, and cost functions and slack variables were optimized using data from the ADNI database (314 cases, 386 healthy controls). The probability of AD is computed by BAAD from the set of all regions of interest and is shown as an AD score (ADS). The accuracy and post-diagnostic odds ratio using BAAD AD scores were assessed at 89.6% and 134.1, respectively. We used the AIBL database (72 AD cases, 447 healthy controls) as an application phase for validation by comparison to results from VSRAD (voxel-based specific regional analysis system for AD) software. The accuracy and the post-diagnostic odds ratio for AD scores were 86.1% and 47.9 for BAAD but 84.8% and 14.9 for VSRAD. This suggests that the BAAD approach more fully exploits the potential of structural analysis to support AD diagnosis.