Augmenting radiological assessment of imaging evident dementias with radiomic analysis

Accurate differential diagnosis of dementia is essential for guiding timely treatment, particularly as anti-amyloid therapies become more widely available and require precise patient characterization. Here, we developed a radiomics-based machine learning (ML) approach to enhance neuroimaging assessments in distinguishing Alzheimer's disease (AD) from other imaging-evident dementias (OIED). We retrospectively analyzed 1041 individuals from the National Alzheimer's Coordinating Center with confirmed dementia diagnoses and at least one T1 or T2/FLAIR MRI scan. Using FastSurfer and a Lesion Prediction Algorithm, we extracted volumetric and lesion features, which were then used to train ML models. Model performance was compared to the independent evaluations of seven fellowship-trained neuroradiologists. The classifier achieved an AUROC of 0.79 ± 0.01 for AD and 0.66 ± 0.03 for OIED, performing comparably to expert assessments. Interpretation using SHAP values showed strong alignment with imaging features known to align with AD or OIED, respectively. These findings highlight the potential of radiomics to augment neuroimaging workflows.