Explainable AI for MRI Alzheimer's disease classification: A comparative analysis

We introduce a quantitative pipeline for region-level explanations of an Alzheimer's prediction model using four post hoc explainable AI (XAI) methods (DeepLIFT, Layer Gradient × Activation, Occlusion sensitivity analysis, and XGrad-CAM). Explanations are mapped to anatomical regions of interest using Freesurfer and summarized as the percentage of volume per region that is deemed relevant by the XAI method. We validate the pipeline in a cross-sectional ADNI cohort, where the explanations (except XGrad-CAM) converged on hippocampus/amygdala as most discriminative for CN, whereas relevant regions for AD were method-dependent. In longitudinal data from subjects converting from cognitively normal to Alzheimer's disease, the relevance of caudate/putamen/pallidum increased across most XAI methods, with FDR-corrected significant trends for three of four methods. In amyloid-positive MCI subjects, attribution profiles were intermediate between AD and CN groups. We additionally evaluated XAI fidelity using perturbation analyses and quantified cross-modal correspondence by correlating regional MRI relevance with amyloid and tau positron emission tomography (PET) standardized uptake value ratios. We observed weak correlations between XAI maps and PET across cortical and subcortical regions, suggesting that MRI and PET provide complementary information. Our findings show that XAI results are method- and class-dependent for AD classification. Still, all methods might convey relevant complementary information that can be used by end users to better understand the biological processes behind the patterns used to classify AD subjects. Thus, we suggest applying different XAI methods for AD classification.