Spatio-temporal convolution for classification of alzheimer disease and mild cognitive impairment

Dementia refers to the loss of memory and other cognitive abilities. Alzheimer's disease (AD), which patients eventually die from, is the most common cause of dementia. In USA, %60 to %80 of dementia cases, are caused by AD. An estimate of 5.2 million people from all age groups have been diagnosed with AD in 2014. Mild cognitive impairment (MCI) is a preliminary stage of dementia with noticeable changes in patient's cognitive abilities. Individuals, who bear MCI symptoms, are prone to developing AD. Therefore, identification of MCI patients is very critical for a plausible treatment before it reaches to AD, the irreversible stage of this neurodegenerative disease.

Development of machine learning algorithms have recently gained a significant pace in early diagnosis of Alzheimer's disease (AD). In this study, a (2+1)D convolutional neural network (CNN) architecture has been proposed to distinguish mild cognitive impairment (MCI) from AD, based on structural magnetic resonance imaging (MRI). MRI scans of AD and MCI subjects were procured from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database. 507 scans of 223 AD patients and 507 scans of 204 MCI patients were obtained for the computational experiments.

The outcome and robustness of 2D convolutions, 3D convolutions and (2+1)D convolutions were compared. The CNN algorithms incorporated 2 to 6 convolutional layers, depending on the architecture, followed by 4 pooling layers and 3 fully connected layers. (2+1)D convolutional neural network model resulted in the best classification performance with 85% auc score, in addition to an almost two times faster convergence compared to classical 3D CNN methods.

Application of (2+1)D CNN algorithm to large datasets and deeper neural network models can provide a significant advantage in speed, due to its architecture handling images in spatial and temporal dimensions separately.