Structural imaging investigations commonly apply a segmentation step followed by the extraction of feature data that can be used to compare or discriminate groups. We present a framework for such a study based on automated multi-atlas segmentation followed by the extraction of low-level morphological features, volumes and overlaps, for classification. A spectral analysis step is used to transform pairwise overlap information into feature data that relate to individual subjects. Applying the framework to a group of controls and patients with mild dementia, we compare the volume- and overlap-based classification performance using both supervised and unsupervised classifiers. The results indicate that unsupervised classification following a spectral analysis of label overlaps performs very well, outperforming classifiers that use volumes alone.