Development and validation of an amyloid PET dementia of the Alzheimer's type (DAT) score

Abstract

Background Alzheimer's disease (AD) can be debilitating if left untreated, but its progression may be altered through early detection. Objective To develop and evaluate a convolutional neural network (CNN) for detecting AD from amyloid PET brain images and to investigate the regions contributing to model predictions. Methods A 3D CNN with residual connections was developed to classify amyloid PET brain volumes. Amyloid PET data were obtained from the Alzheimer's Disease Neuroimaging Initiative (ADNI), with approximately 600 images from cognitively normal control (NC) and dementia of the Alzheimer's type (DAT) participants used for training, validation, and testing. Performance was assessed using repeated 5-fold cross-validation (10 total folds). The model was also evaluated across the AD continuum, including unstable normal control (uNC), progressive normal control (pNC), stable mild cognitive impairment (sMCI), progressive mild cognitive impairment (pMCI), and early DAT (eDAT). Saliency and class activation maps were generated to identify regions contributing to predictions. Results The model achieved a mean testing accuracy of 92% across the 10 folds. Across the disease continuum, accuracies were 76% for uNC, 78% for sMCI, 24% for pNC, 65% for pMCI, and 78% for eDAT. Saliency and class activation maps highlighted the putamen, thalamus, hippocampus, corpus callosum, and posterior cingulate cortex, regions previously implicated in AD pathology. Conclusions The proposed 3D CNN accurately distinguished DAT from cognitively normal controls using amyloid PET imaging and showed promising performance across the AD continuum. Model interpretation identified biologically relevant brain regions, supporting the potential of deep learning for early AD detection and clinical decision support.