Inclusion of intracranial volume as a covariate feature improves MRI-based Alzheimer's disease classification

Structural MRI-based regional volumes are widely used for Alzheimer's disease (AD) classification, but inter-individual variability in intracranial volume (ICV) introduces confounding. Traditional adjustment methods use region-of-interest (ROI)/ICV ratios or residual adjustment during pre-processing, yet no consensus exists on the optimal method. This study tests whether explicitly including ICV as a covariate (ROI + ICV) improves classification compared with ratio, residual adjustment, and the unadjusted baseline.

T1-weighted MRIs from ADNI1 (n = 1423) and MIRIAD (n = 69) were processed with FreeSurfer to extract eight AD-related ROI volumes and ICV. Four feature configurations (ROI-only, ROI/ICV, Residual ROI, ROI + ICV) were benchmarked across six classifiers for cognitive normal (CN)-AD, CN-mild cognitive impairment (MCI), and MCI-AD. Performance was assessed with AUROC and F1 using Friedman and post hoc tests. In addition, feature attribution was examined with permutation importance and SHAP.

ROI + ICV consistently produced the largest performance gains over ROI-only in CN-AD and CN-MCI, outperforming ratio and residual adjustment across most classifiers. These improvements generalized to the independent MIRIAD dataset. SHAP analyses showed that the directional effect of ICV reversed across strategies: under ratio or residual adjustment, larger ICV decreased AD probability, whereas in ROI + ICV, larger ICV increased it. This highlights ICV's contextual influence on model decisions.

Pre-processing-based adjustments do not fully remove ICV effects and can distort ROI-ICV relationships. Explicit covariate inclusion avoids these issues and yields more consistent, generalizable improvements. Thus, ICV should be modeled rather than removed, making ROI + ICV the preferred default ICV-handling strategy for MRI-based AD classification.