Predicting Conversion to Mild Cognitive Impairment in Cognitively Normal with Incomplete Multi-modal Neuroimages

Assessing clinical progression from cognitively normal (CN) to mild cognitive impairment (MCI) is crucial for early intervention before the onset of cognitive decline. Multi-modal neuroimaging data has provided supplementary biomarkers for computer-aided prediction of neurodegeneration diseases. However, it is still unknown whether tau uptake in positron emission tomography (PET) provides much power for identifying progressive CN who will convert to MCI, since subjects usually lack tau PET scans. In this study, we proposed a neuroimage synthesis network to impute missing tau PET images based on their corresponding T1-weighted magnetic resonance imaging (MRI) scans. With the real MRI and synthetic PET data after imputation, we applied support vector machine classifiers on regional measurement of anatomical features extracted from pre-defined atlases for prediction. Experimental results on Alzheimer's Disease Neuroimaging Initiative dataset suggest that our neuroimage synthesis network synthesized reasonable neuroimages and complementary information provided by tau PET improved the accuracy of identification.