TY - GEN
T1 - Groupwise segmentation improves neuroimaging classification accuracy
AU - Wang, Yaping
AU - Jia, Hongjun
AU - Yap, Pew Thian
AU - Cheng, Bo
AU - Wee, Chong Yaw
AU - Guo, Lei
AU - Shen, Dinggang
PY - 2012
Y1 - 2012
N2 - 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.
AB - 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.
UR - http://www.scopus.com/inward/record.url?scp=84868237036&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=84868237036&partnerID=8YFLogxK
U2 - 10.1007/978-3-642-33530-3_16
DO - 10.1007/978-3-642-33530-3_16
M3 - Conference contribution
AN - SCOPUS:84868237036
SN - 9783642335297
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 185
EP - 193
BT - Multimodal Brain Image Analysis - Second International Workshop, MBIA 2012, Held in Conjunction with MICCAI 2012, Proceedings
T2 - 2nd International Workshop on Multimodal Brain Image Analysis, MBIA 2012, Held in Conjunction with the 15th International Conference on Medical Image Computing and Computer Assisted Intervention, MICCAI 2012
Y2 - 1 October 2012 through 5 October 2012
ER -