Search

Alzheimer’s disease (AD) is an irreversible and progressive disease that begins with mild memory loss possibly leading to loss of the ability to respond to the environment. Diffusion-weighted magnetic resonance images (DW-MRI)-derived brain network has been widely used as an important tool to study AD. Many methods have been proposed to reconstruct brain structural networks in the literature. However, due to the variance of each method, the resulted brain networks are distinct. Consequently, the downstream analyses and results are varied, which significantly affect the clinical implications of the conclusions. Moreover, due to the limitation of the available imaging data, it may also be a challenge to effectively capture highly discriminative features related to AD, which can further affect the predictive model’s performance. In this thesis, we address these challenges by proposing a new CNN (Convolution Neural Network) and LSTM (Long Short-Term Memory) mixed model for AD classification. Here we treat different networks from the same subject as multi-view data and apply our new model on two independent publicly available AD cohorts (202 subjects from the 2nd stage of Alzheimer’s disease neuroimaging initiative or ADNI2 and 445 subjects from National Alzheimer’s Coordinating Center or NACC). ADNI2 has four group data (normal control or NC, early mild cognitive impairment or EMCI, late mild cognitive impairment or LMCI and AD) while NACC has only three group data (NC, MCI, and AD). Our experimental results show that our new model can achieve around 96% accuracy for multi-group classifications for both cohorts, significantly outperformed those baseline methods whose accuracy is around 75%.