Groupwise segmentation improves neuroimaging classification accuracy

Accurate diagnosis of Alzheimer's disease (AD), especially mild cognitive impairment (MCI), is critical for treatment of the disease. Many algorithms have been proposed to improve classification performance. While most existing methods focus on exploring different feature extraction and selection techniques, in this paper, we show that the pre-processing steps for MRI scans, i.e., registration and segmentation, significantly affect the classification performance. Specifically, we evaluate the classification performance given by a multi-atlas based multi-image segmentation (MABMIS) method, with respect to more conventional segmentation methods. By incorporating tree-based groupwise registration and iterative groupwise segmentation strategies, MABMIS attains more accurate and consistent segmentation results compared with the conventional methods that do not take into account the inherent distribution of images under segmentation. This increased segmentation accuracy will benefit classification by minimizing errors that are propagated to the subsequent analysis steps. Experimental results indicate that MABMIS achieves better performance when compared with the conventional methods in the following classification tasks using the ADNI dataset: AD vs. MCI (accuracy: 71.8%), AD vs. healthy control (HC) (89.1%), progressive MCI vs. HC (84.4%), and progressive MCI vs. stable MCI (70.0%). These results show that pre-processing the images accurately is critical for neuroimaging classification.