Multicentre study, Diagnostic or prognostic study, Retrospective, Dementia, Computer Applications-Detection, diagnosis, MR, CAD, Neuroradiology brain, Artificial Intelligence, Artificial Intelligence and Machine Learning
S. Alam1, A. S. Lundervold2, A. Lundervold2; 1Bergen, Hordaland, Norway/NO, 2Bergen/NO
Alzheimer's disease (AD) is a progressive neurodegenerative disorder associated with damage of brain cells and brain shrinkage . Automatic classification of AD or mild cognitive impairment (MCI) from cognitively normal (CN) subjects using magnetic resonance imaging (MRI) requires a good representation of brain features and trained classifiers.
Brain shrinkage features relating to AD can be assessed with MRI by using an atlas providing prior knowledge of the structure of the targeted brain regions.
It has been demonstrated that using a multi-atlas approach achieves better feature representation than using a single atlas only . Multi-atlas-based likelihood fusion (MALF) is a hierarchical method unifying skull stripping and extraction of volumes (features) of subcortical regions of interest (ROIs) . It provides better segmentation accuracy both for skull stripping and segmentation of subcortical regions compared to other methods tested in . The pairwise correlation between ROI volumes across a set of brain regions (i.e., ROI-C features) can potentially provide additional valuable information. To test this hypothesis, we employed a multi-kernel support vector machine (SVM), a model that addresses both linearly and nonlinearly separated data. By using multiple kernels, the dependency of kernel parameters is reduced.
More specifically, we investigated (i) whether classification performance differs when using a set of single ROI features versus a set of ROI-correlative (ROI-C) features, or a combined set of single ROI and ROI-C features, and (ii) the ranking of the combined (ROI+ROI-C) features based on their importance for detecting AD versus CN, AD versus MCI, and CN versus MCI.