Ensembles of Patch-Based Classifiers for Diagnosis of Alzheimer Diseases

There is ongoing research for the automatic diagnosis of Alzheimer's disease (AD) based on traditional machine learning techniques, and deep learning-based approaches are becoming a popular choice for AD diagnosis. The state-of-the-art techniques that consider multimodal diagnosis have been shown to have accuracy better than a manual diagnosis. However, collecting data from different modalities is time-consuming and expensive, and some modalities may have radioactive side effects. Our study is confined to structural magnetic resonance imaging (sMRI). The objectives of our attempt are as follows: 1) to increase the accuracy level that is comparable to the state-of-the-art methods; 2) to overcome the overfitting problem, and; 3) to analyze proven landmarks of the brain that provide discernible features for AD diagnosis. Here, we focused specifically on both the left and right hippocampus areas. To achieve the objectives, at first, we incorporate ensembles of simple convolutional neural networks (CNNs) as feature extractors and softmax cross-entropy as the classifier. Then, considering the scarcity of data, we deployed a patch-based approach. We have performed our experiment on the Gwangju Alzheimer's and Related Dementia (GARD) cohort dataset prepared by the National Research Center for Dementia (GARD), Gwangju, South Korea. We manually localized the left and right hippocampus and fed three view patches (TVPs) to the CNN after the preprocessing steps. We achieve 90.05% accuracy. We have compared our model with the state-of-the-art methods on the same dataset they have used and found our result comparable.