Keywords: Alzheimer's disease, imaging, statistical learning, biomarkers, predictive modeling
Neuroimaging analysis is shifting away from univariate, voxel-wise methods in favor of more flexible multivariate models. Supervised and unsupervised statistical learning methods are being applied to complex, high-dimensional neuroimaging data with the promise of better understanding of disease processes in the brain and more powerful biomarkers for disease diagnosis. The application of these methods at different stages in the image processing and analysis pipeline has generated a multitude of open statistical problems. Some of the challenges in the statistical analysis of neuroimaging data include assessing data quality, image registration, intensity normalization, feature extraction, estimation and interpretation of image-based biomarkers, generalizability of findings, and accounting for the observational nature of neuroimaging studies. After a broad overview of these issues, specific attention will be given to addressing confounding in image-based biomarkers for Alzheimer’s disease.