Abstract

Independent component analysis (ICA) is widely used for extracting brain functional network (FN) from fMRI data, but the low signal-to-noise ratio in fMRI data makes FNs contain a lot of noise. Previous methods that remove the noise-related components obtained from ICA to reconstruct fMRI largely depend on the trained models and fail to directly optimize FN. Here, based on our previously proposed group information guided ICA (GIG-ICA), we incorporate a smoothness term to construct a multi-objective function so as to effectively remove the noises in FNs in addition to guaranteeing the independence and correspondence of FNs. Importantly, different kinds of noises can be handled independently or jointly, adaptively or knowledge-based. We validate our method using both the simulated and real fMRI data. Using the simulated data added with different noises, the spatial accuracy of FNs obtained by our method (named GIG-sICA) is higher than that obtained by the original GIG-ICA. Using fMRI data of 134 healthy controls and 123 schizophrenia patients, our method yields greater group differences in most FNs compared with GIG-ICA. In summary, our method is effective in extracting smooth and accurate brain FNs with less noise, and helpful to identify stable biomarkers and predict mental illnesses.