Introduction1.1. Alzheimer’s DiseaseAlzheimer’sdisease is a brain disorder that progresses over time. It most commonly happens in patients’ latestage life (Burns& Iliffe, 2009). It involves the death of nerve cells and thus leads to the reduction ofbrain volume (Fig.
1), which impactbrain functions. The most noticeable early stage symptomincludes short-term memory loss. As thedisease progress, patients may show disturbances in language, orientation,reasoning and perception (Weiner et al.,2010).
At the same time,both the annual incidence of Alzheimer’s disease and the cost of caring orAlzheimer’s disease are on the rise. Itis believed that the number of Alzheimer’s disease patients will be doubled by2050. The estimated amount of money Americansspent on Alzheimer’s disease treatment and caring is $259 billion USD in 2017. (Alzheimer’sAssociation, 2015) Fig.1The figure shows a Alzheimer’s disease brain comparingto normal brain. The shrinkage of wholebrain size, particularly Hippocampus as well as the shriveled cortex can beseen. 1.
2. TraditionalAlzheimer’s Disease DiagnosisTraditional Alzheimer’s disease is diagnosed clinicallybased on the person’s description, medical records and behaviors (Mendez, 2006). Doctors cannot guarantee an acceptable accuracy for diagnosis withoutclose examine of the brain tissue under a microscope, which can only be doneafter the patient’s death. 1.3.
Machine Learning in Alzheimer’sDisease DiagnosisMachinelearning enables computers to learn without specifically programmedinstructions and guidelines (Samuel, 2000). In the field of Alzheimer’s diseasediagnosis, the aim is to increase data processing ability and decrease the specialtydomain expert knowledge needed to detect brain abnormality at a high accuracy (Plis et al., 2014). Many studies were done to develop suchmodels. 1.4. Prior ResearchFeatureslike Brain volume, hippocampus shape and cortex character can be used to diagnoseAlzheimer’s disease utilizing several types of brain images.
Structural MRI and function MRI are among themost popular. Priorresearchers started with segmenting brains to different part, extractingvoxel-wise features. These features arerepresented by vectors of multi-scalar quantifications. Fan successfully segmented the brain to greymatter, white matter and computed their corresponding voxel-wise densities (Fan, Shen, Gur, Gur, & Davatzikos, 2007). These values were later on converted to vector form that can be used forclassification. Further on top of theirwork, Lerch et al computed cortical thickness by measuring the gap betweenselected points at grey matter and white matter (Lerch et al., 2008). ConvolutionalNeural Networks were also used for the purpose of Alzheimer’s disease diagnosisin mainly two directions: analyzing medical record and analyzingbio-images.
Aston and Gunn first proposedConvolutional Neural Network in 2D dimension to extract unique features fromMRI slices. And later, Payan employed a3D Convolutional Neural Network for this purpose (Payan & Montana, 2015). Recent work by Hosseini-Asl implemented a 3DAuto-encoder at the lower layer. It ispre-trained by a set of dementia data and the upper fully connected layers willbe used for fine tuning by specific data domain (Hosseini-Asl et al.