Deep learning combining FDG-PET and neurocognitive data accurately predicts MCI conversion to Alzheimer's dementia 3-year post MCI diagnosis

Introduction: This study reports a novel deep learning approach to predict mild cognitive impairment (MCI) conversion to Alzheimer's dementia (AD) within three years using whole-brain fluorodeoxyglucose (FDG) positron emission tomography (PET) and cognitive scores (CS). Methods: This analysis consisted of 150 normal controls (CN), 257 MCI, and 205 AD subjects from ADNI. FDGPET and CS were obtained at MCI diagnosis to predict AD conversion within three years of MCI diagnosis using convolutional neural networks. Results: Neurocognitive scores predicted better than FDG-PET per se, but the best model was a combination of FDG-PET, age, and neurocognitive data, yielding an AUC of 0.785 +/- 0.096 and a balanced accuracy of 0.733 +/- 0.098. Saliency maps highlighted putamen, thalamus, inferior frontal gyrus, parietal operculum, precuneus cortices, calcarine cortices, temporal gyrus, and planum temporale to be important for prediction. Discussion: Deep learning accurately predicts MCI conversion to AD and provides neural correlates of brain regions associated with AD conversion.