Characterizing brain imaging features associated with ADAS-Cog13 sub-scores with 3D convolutional neural networks

To date, few advanced machine learning models have been developed for investigating the associations between features from brain imaging and individual Alzheimers disease (AD) related cognitive functional changes. Additionally, how these associations differ among different imaging modalities is unclear. Here we investigated 3D convolutional neural network (CNN) models which were trained to predict sub-scores in 13-item Alzheimers Disease Assessment Scale - Cognitive Subscale (ADAS-Cog13) based on MRI and FDG-PET brain imaging data obtained from the ADNI database. We found that each key ADAS-Cog13 sub-score was associated with a specific set of brain features within an imaging modality. Overall, sub-scores were strongly associated with structural changes of subcortical regions including amygdala, hippocampus, and putamen, and were associated with metabolic changes of cortical regions including the cingulated gyrus, occipital cortex, middle front gyrus, precuneus cortex, and the cerebellum. Our findings provided insights into complex AD etiology. Our analytical pipeline can also be utilized to study other brain diseases.