Data Representation and Classification of Alzheimer’s Disease

Application of machine learning algorithms to information of magnetic resonance imaging (MRI) is a widespread approach to differentiate Alzheimer’s disease (AD) patients and healthy controls (HC). Since a variety of brain representations are used by different studies, it is necessary that the influence of the chosen brain atlas on the model performance is investigated. Therefore, the goal is to analyse the effect which is caused by varying granularity of an atlas. In addition, to find acceptance in the medical community, the model must be able to identify biologically relevant regions. Thereby, it can be ensured that the model will reliably identify patients in future applications and is not based on sample-specific characteristics. For this reason, the regions selected for classification by support vector machine (SVM) to differentiate AD vs HC are analysed. Lastly, features that are not selected by a given model are generally disregarded. Since those features could potentially still contain relevant information, they are examined in this study. Different granularities of the Schaefer atlas, with parcellations ranging from 173 to 1273 parcels, were used to extract features from structural images of AD patients and healthy controls. Subsequently, SVM classifiers were trained on the features derived from the different parcellations and their influence was evaluated based on the performance of the resulting model. Biological relevance of the selected features was verified by confirming their role in AD with current literature. Non-selected features were singled out and used to train a non-selected feature model (NFM). Relevance of the non-selected features was evaluated based on performance of the NFM. Evaluation of the obtained accuracies showed that the granularity of the atlas affects the model performance on 1.5 Tesla images of AD patients and HC. Accuracies ranged from 87% for the 173 parcel parcellation, to 83% for the 1273 parcel parcellation. Classification of 3 Tesla images was not significantly affected, with all models achieving accuracies around 91%. Biological relevance of the selected features could be confirmed by literature, although it was evident that not all relevant regions were included in the model. Examination of the NFM revealed that a model based on non-selected features could still classify AD vs HC with an accuracy of 76%. The findings suggest that future atlas-based approaches should pay more attention to the effect of the selected atlas. In addition, the ability of SVM to select biologically relevant regions supports its implementation for diagnosis of AD in the clinic. Lastly, the results indicate that investigation of non-selected features could provide additional insight into the relevance of certain regions for the studied disease