| Summary: | The accurate diagnosis and prediction for individuals is crucial in computer-aided diagnosis of Alzheimer's disease (AD). The existing structural magnetic resonance imaging based classification methods of AD diagnosis mainly focus on the voxel level, region level and patch level morphological pattern analysis. However, most of these methods extract features with high dimension which may lead to overfitting problem. Besides, the interaction of different patches is not considered in the classifier ensemble. In this article, we propose a novel anatomical landmarks and directed acyclic graph (DAG) network feature learning based classification algorithm for the diagnosis of AD individuals. First, the anatomical feature patches of gray matter image are identified by the morphological and statistical analysis. Second, a simple and efficient DAG convolutional neural network is proposed to extract the discriminative deep features of image representation. Especially, the deep features are obtained by fusing feature maps of different network levels which contain semantic high-level and high-resolution low-level features. Finally, support vector machine and deep features are utilized to construct the classification model and predict the individual of AD. Experiments on three public datasets including ADNI-1, ADNI-2 and MIRIAD demonstrate that the proposed method can effectively improve the classification performance compared with the state-of-the-art methods for AD diagnosis.
|
|---|
