Hierarchical ensemble learning for Alzheimer's disease classification

In this paper, we propose to tackle the problem of Alzheimer's Disease (AD) classification by a novel Hierarchical Ensemble Learning (HEL) framework. Given an MRI image of a subject, our method will divide it into multiple slices, and generate the classification result in a coarse-to-fine way: First, for each slice, multiple pre-trained deep neural networks are adopted to extract features, and classiflers trained with each type of these features are used to generate the coarse predictions; Second, we employ ensemble learning on the coarse results to generate a refined result for each slice; At last, the given subject is classified based on the refined results aggregated from all slices. Using pre-trained networks for feature extraction can reduce the computational costs of training significantly, and the ensemble of multiple features and predicted results from slices can increase the classification accuracy effectively. Hence, our method can achieve the balance between efficiency and effectiveness. Experimental results show that the HEL framework can obtain notable performance gains with respect to various features and classifiers.