The effect of groupness constraint on the sensitivity and specificity of canonical correlation analysis, a multi-modal anatomical and functional MRI study

• Published in: Biomedical signal processing and control Vol. 68; p. 102698
• Main Authors: Mohammadi-Nejad, Ali-Reza, Hossein-Zadeh, Gholam-Ali, Shahsavand Ananloo, Esmaeil, Soltanian-Zadeh, Hamid
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.07.2021
• Subjects: Canonical correlation analysis (CCA); Group-Structured; Magnetic resonance imaging (MRI); Multi-Modal data fusion; Overlapping and disjoint groups; Schizophrenia; Magnetic resonance imaging (MRI); Schizophrenia; Overlapping and disjoint groups; Group-Structured; Multi-Modal data fusion; Canonical correlation analysis (CCA)
• ISSN: 1746-8094; 1746-8108
• DOI: 10.1016/j.bspc.2021.102698

•A group-structured sparse CCA method for the fusion of two neuroimaging modalities.•The approach allows considering disjoint and overlapping groups in the data fusion.•The gssCCA is general and includes the existing CCA-based methods as special cases.•The overlapping gssCCA method could better recover the true correlations. While neuroimaging studies reveal that several brain regions may participate in multiple groups (networks), this group overlap is neglected in multi-modal data fusion frameworks. Indeed, it is not clear how much “information” is lost due to this negligence. To study this issue, we present a group-structured sparse canonical correlation analysis (gssCCA) technique by utilizing groupness and sparsity constraints in a unified fusion framework. The approach allows to: 1) compare structures of disjoint and overlapping groups (networks); and 2) consider appropriate levels of overlap among groups. We use simulations to investigate the performance of the proposed approach and compare overlapping gssCCA, disjoint gssCCA, and ssCCA. The gains of considering an overlapping groupness constraint are significant since it can increase detection sensitivity and specificity of associations between multi-modal datasets. They also demonstrate that, even when the ROIs are assumed to be disjoint, the lost structural information is less than the conditions that we do not have any groupness information. We also apply the methods to experimental anatomical magnetic resonance imaging (MRI) and resting-state functional MRI (rs-fMRI) data of schizophrenia (SZ) patients and control subjects (CS). The results show that the first pair of canonical variates (CVs) capture better classification accuracy between SZ and CS with a correlation of 88 % (p-value of less than 1 × 10−6). The extracted CVs show the most correlated clusters between anatomical and functional datasets. These clusters explain differences between the two groups in their own modality that are maximally correlated with the corresponding clusters in the other modality.