Development and validation of machine learning method for an Amyloid-PET dementia of Alzheimer’s type (DAT) score

Abstract
Amyloid positron emission tomography (amyloid-PET) is a non-invasive, in-vivo, tracer-specific neuroimaging that detects amyloid plaques in the brain. Studies have confirmed that 18F-florbetapir (AV45) is one of the amyloid traces that can indicate the development of dementia of Alzheimer’s type (DAT). Therefore, machine learning methods can use the images for early detection of DAT. This thesis proposes developing, validating, and testing the three-dimensional convolutional neural network (3D CNN) method to produce a score indicating the subject’s possibility of developing DAT. AV45-PET images were processed, co-registered with their corresponding magnetic resonance imaging (MRI), and segmented to hold subcortical segmentation (aseg) and cortical parcellation (aparc) segmentation be-fore being used in the model. To better understand the model, the subjects were divided into seven subgroups: sNC (stable normal control), uNC (unstable NC), pNC (progressive NC), sMCI (stable mild cognitive impairment), pMCI, eDAT (converted to DAT in ADNI window), and sDAT (joined study with DAT diagnosed). The 3D CNN was trained and val-idated using the sDAT and sNC images while testing was done on the remaining subgroups. The validation averaged results were, 84.9% precision, 66.2% sensitivity, and 0.612 AUC. The model diagnosed the eDAT, pMCI, pNC, sMCI, and uNC subjects with accuracy of 74.7%, 56.1%, 29.41%, 80%, and 85.56% respectively.