Abstract
In blind source separation (BSS) for multisubject functional magnetic resonance imaging (fMRI) data, dimensionality reduction is generally performed for multiple times. This leads to the challenge of determining the number of the retained dimensionality, i.e. the order of BSS models, which dramatically influences the validity and performance of BSS models. In this study, a multisubject analysis method robust to order selection is developed. This approach remains effective for slight dimensionality reduction and thus utilizes more information from original data. Inspired by the idea of signal-intensity-maximization technology, which can suppress the overfitting that occurs during insufficient dimensionality reduction, we rotate the reduced dimensions to the optimized direction so that rotated components have the most significant intensity and smoothness. Because the optimized dimensions contain more useful information, involving dimensional optimization stage can reduce the negative impact of dimensionality reduction in multisubject data analysis. The experiments on simulated data and real fMRI data showed that involving dimensional optimization improves the validity and performance of the BSS model in analyzing multisubject data. The proposed method works better across a wide range of dimensionality reduction levels, allows inaccurate order selection, maintains more useful information, and is suitable for multisubject fMRI analysis, which requires multiple dimensionality reduction.
Highlights
Blind source separation (BSS) for functional magnetic resonance imaging data is an important topic in biomedical image analysis [1]–[3]
COMPARISON ON SIMULATED DATA On the simulated data, to evaluate the role of dimensional optimization, our method was firstly compared with three simplified procedures
Among the group components separated by each algorithm, the source that had the highest correlation with the simulated source was selected as the output of the algorithm
Summary
Blind source separation (BSS) for functional magnetic resonance imaging (fMRI) data is an important topic in biomedical image analysis [1]–[3]. When analyzing multi-set data, classical BSS methods meet the challenge of keeping the coherence among the estimated sources from multiple datasets [4]. Much effort has been made to achieve BSS for multisubject fMRI data. Some researchers reorganize data so that the multisubject data can be analyzed by classical BSS methods. Calhoun et al concatenated multiple datasets into a group prior to BSS, and proposed group independent component analysis (group ICA) [5]. Some researchers extend classical BSS models so that multi-set data can be directly analyzed. Used methods in this scope include independent vector analysis (IVA) [6]–[8], multi-set canonical correlation
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have