TY - GEN
T1 - Learning-based meta-algorithm for MRI brain extraction
AU - Shi, Feng
AU - Wang, Li
AU - Gilmore, John H.
AU - Lin, Weili
AU - Shen, Dinggang
PY - 2011
Y1 - 2011
N2 - Multiple-segmentation-and-fusion method has been widely used for brain extraction, tissue segmentation, and region of interest (ROI) localization. However, such studies are hindered in practice by their computational complexity, mainly coming from the steps of template selection and template-to-subject nonlinear registration. In this study, we address these two issues and propose a novel learning-based meta-algorithm for MRI brain extraction. Specifically, we first use exemplars to represent the entire template library, and assign the most similar exemplar to the test subject. Second, a meta-algorithm combining two existing brain extraction algorithms (BET and BSE) is proposed to conduct multiple extractions directly on test subject. Effective parameter settings for the meta-algorithm are learned from the training data and propagated to subject through exemplars. We further develop a level-set based fusion method to combine multiple candidate extractions together with a closed smooth surface, for obtaining the final result. Experimental results show that, with only a small portion of subjects for training, the proposed method is able to produce more accurate and robust brain extraction results, at Jaccard Index of 0.956±0.010 on total 340 subjects under 6-fold cross validation, compared to those by the BET and BSE even using their best parameter combinations.
AB - Multiple-segmentation-and-fusion method has been widely used for brain extraction, tissue segmentation, and region of interest (ROI) localization. However, such studies are hindered in practice by their computational complexity, mainly coming from the steps of template selection and template-to-subject nonlinear registration. In this study, we address these two issues and propose a novel learning-based meta-algorithm for MRI brain extraction. Specifically, we first use exemplars to represent the entire template library, and assign the most similar exemplar to the test subject. Second, a meta-algorithm combining two existing brain extraction algorithms (BET and BSE) is proposed to conduct multiple extractions directly on test subject. Effective parameter settings for the meta-algorithm are learned from the training data and propagated to subject through exemplars. We further develop a level-set based fusion method to combine multiple candidate extractions together with a closed smooth surface, for obtaining the final result. Experimental results show that, with only a small portion of subjects for training, the proposed method is able to produce more accurate and robust brain extraction results, at Jaccard Index of 0.956±0.010 on total 340 subjects under 6-fold cross validation, compared to those by the BET and BSE even using their best parameter combinations.
UR - http://www.scopus.com/inward/record.url?scp=80053493942&partnerID=8YFLogxK
U2 - 10.1007/978-3-642-23626-6_39
DO - 10.1007/978-3-642-23626-6_39
M3 - Conference contribution
C2 - 22003714
AN - SCOPUS:80053493942
SN - 9783642236259
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 313
EP - 321
BT - Medical Image Computing and Computer-Assisted Intervention, MICCAI 2011 - 14th International Conference, Proceedings
T2 - 14th International Conference on Medical Image Computing and Computer Assisted Intervention, MICCAI 2011
Y2 - 18 September 2011 through 22 September 2011
ER -