<p>Alzheimer’s disease is a complex neurodegenerative disorder and the leading cause of dementia worldwide. Learning-based techniques applied to magnetic resonance imaging (MRI) have recently shown strong potential for automated diagnosis. Accurate classification typically relies on high-resolution (HR) 3D MRI acquired with thin axial slices to reduce partial-volume artefacts, capture fine anatomical details, and improve diagnostic performance. However, acquiring such data is time-consuming, costly, and prone to motion artefacts and patient discomfort. Super-resolution methods offer a promising alternative by reconstructing HR 3D images from lower-resolution scans and enabling shorter acquisition times. In this study, we propose a novel pipeline that applies super-resolution to through-plane undersampled 3D magnetic resonance images and demonstrates that the resulting volumes preserve Alzheimer’s disease diagnostic accuracy comparable to that achieved using fully sampled HR scans. We compare different state-of-the-art super-resolution methods from distinct methodological families, with the best-performing method achieving an F1 score of 65.6, close to the HR reference of 65.7 and substantially higher than the low-resolution baseline of 55.9. Furthermore, we investigate whether standard image quality metrics (e.g. pixel-based metrics) are sufficient to assess the contribution of super-resolution to the clinical evaluation of Alzheimer’s disease. To this end, we compare them with machine learning–based measures, such as maximum mean discrepancy, and surface-based metrics derived from segmented anatomical structures, highlighting their limitations in clinically oriented evaluations.</p>