Comparison of three group ICA methods for multi-subject fMRI data analysis

• Published in: International Conference on Information Science and Technology pp. 1276 - 1280
• Main Authors: Mingqi Hui, Li Yao, Zhiying Long
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.03.2011
• Subjects: Correlation; Humans; Independent component analysis; Magnetic resonance imaging; Noise level; Principal component analysis; Tensile stress
• ISBN: 1424494400; 9781424494408
• ISSN: 2164-4357
• DOI: 10.1109/ICIST.2011.5765072

Since spatial independent component analysis (sICA) was introduced to fMRI data analysis of single subject by Mckeown (1998), several group ICA approaches including subject-specific ICA, across-subject averaging, temporal/spatial concatenation and tensor probabilistic ICA (PICA) have been proposed to generate group inference over a group of subjects. By far, the subject-specific ICA, temporal concatenation method and tensor PICA have been applied to fMRI data analysis. Among the three methods, both temporal concatenation method and tensor PICA are the most widely used. However, there hasn't been any comparison of subject-specific ICA, temporal concatenation method and tensor PICA. The current study aims at comparing the three group ICA methods at various noise levels. Simulated experiment based on human resting fMRI data revealed that tensor PICA provided the best overall performance due to the highest spatial detection power and relatively more accurate estimation of time course. Temporal concatenation method provided moderate performance in terms of relatively higher spatial detection power and simpler algorithm. Subject-specific method showed the worst spatial detection power and was the most time consuming in component selection.