Feature extraction by orthonormal projective non‐negative matrix factorization and longitudinal association between brain atrophy patterns and Alzheimer’s disease

Abstract

AbstractBackgroundOrthonormal projective non‐negative matrix factorization (OPNMF) was a recently developed approach for feature extraction techniques of neuroimaging data. However, there has been few studies applying this to brain MRI to better characterize the anatomic changes seen in Alzheimer’s Disease (AD) and correlate these patterns with cognitive changes.MethodWe applied OPNMF with refinement to identify distinct spatial components of voxel‐wise volume loss in the structural MRI (3D T1 MPRAGE images) of subjects with AD (N = 109, mean age = 74.6 (7.9), ADNI2) relative to healthy young controls (N = 104, mean age = 24.0 (2.9), ABIDE1). Non‐negative least squares (NNLS) was then used to extract non‐negative coefficients representing subject‐specific quantitative measures of regional atrophy. This calculation of coefficients was extended to images not used for component discovery which also included Mild Cognitive Impairment (MCI) and controls. To investigate the cross‐sectional and longitudinal associations between cognitive assessments scores and regional atrophy severity in different diagnostic groups, we used multiple linear regression models and liner mixed‐effects models with random intercepts and slopes controlled for age, sex, educational level, APOE ε4 genotypes were utilized. A validation analysis was performed using ADNI3 images.ResultWe derived 6 atrophy components upon the same training set and there was a sharp decrease in the magnitude of the reconstruction error gradient when moving from 5 components to 6 components, and the components have meaningful interpretation. In the whole cohort, atrophy in 5 of the 6 regions was significantly associated with all three cognitive scores. There is evidence of overall positive trends between diagnosis group and atrophy severity changing rate in all 6 regions. The increased atrophy in all 6 regions was associated with decreases in MMSE and MOCA scores and increases in CDRSB scores in the whole cohort as well as in LMCI cohort. From the validation analysis, we showed that our result from OPNMF is stable in atrophy pattern discovery.ConclusionWe developed a robust pipeline for detecting spatially distinct atrophy components in AD and associated these atrophy patterns with cognitive changes which may help us better characterize the underlying pathology of types of AD and other dementias on MRI.