A COPULA LINKED PARALLEL ICA for sMRI and 4D-fMRI fusion
Poster No:
1589
Submission Type:
Abstract Submission
Authors:
Oktay Agcaoglu1, Rogers F. Silva2, Deniz Alacam2, Vince Calhoun3
Institutions:
1Tri-Institutional Center for Translational Research in Neuroimaging and Data Sci, Atlanta, GA, 2TReNDS, Atlanta, GA, 3GSU/GATech/Emory, Decatur, GA
First Author:
Oktay Agcaoglu
Tri-Institutional Center for Translational Research in Neuroimaging and Data Sci
Atlanta, GA
Tri-Institutional Center for Translational Research in Neuroimaging and Data Sci
Atlanta, GA
Co-Author(s):
Introduction:
Current multimodal fusion methods are primarily confined to using second-level 3D fMRI features, such as fALFF, regional homogeneity, or FNC, often neglecting the temporal information in the full 4D fMRI data. To overcome this, we introduce a novel method, Copula Linked Parallel ICA (CLiP-ICA). This method uniquely estimates an unmixing matrix for each modality independently while synchronizing sources through a copula model. We evaluated CLiP-ICA's effectiveness in both a simulation study and a real-world application using fMRI and sMRI data from the Alzheimer's Disease Neuroimaging Initiative (ADNI) study. To our knowledge, this is the inaugural study capable of jointly estimating independent components of fMRI and sMRI while preserving crucial temporal information.
Methods:
Preprocessed fMRI and sMRI datasets are reduced to the dimensions specified by the model order using separate Principal Component Analysis (PCA). CLiP-ICA aims to find two unmixing matrices that maximize the likelihood of a 2D joint distribution, which is implemented as the 1D marginal distribution of each modality and a copula to link them [1]. The marginal distribution can vary between modalities, and the copula can be of any type, such as a Gaussian copula [2]. In this paper, we focus on illustrating the use of Gaussian copulas. During training, batches contain data from corresponding voxels in both modalities
Results:
We generated 4 spatially independent components using a simulation toolbox [3], and later, we used custom MATLAB scripts to generate a second set of 4 components that are linked to the first set with Pearson correlations of 0.05, 0.8, 0.5, and 0.2, while preserving the independency between components and the blob-like structure in the components. Subsequently, we generated two 4x4 mixing matrices and the observed data. To make the simulation more representative of real data scenarios, we ensured that the corresponding rows of the first and second mixing matrices were highly correlated, with correlations of [0.91, 0.96, 0.94, 0.93], resulting in highly similar but not identical mixtures. We then attempted to retrieve sources using both conventional JICA and our proposed CLiP-ICA. CLiP-ICA was able to retrieve the original sources almost perfectly (correlation of 0.99), while JICA only achieved partial retrieval (correlation of 0.53).
We also tested CLiP-ICA on 864 fMRI and sMRI scans from ADNI. The fMRI data were reduced to 30 PCA (45 at the subject level and 30 at the group level), and the gray matter images were also reduced to 30 PCA. Then, we ran CLiP-ICA with a Gaussian copula, where linkage parameters ranged linearly from 0.95 to 0.65. After training, we compared the sMRI and fMRI components. Figure 2 shows the aggregated structural FNC results. Subsequently, utilizing back reconstruction enabled the calculation of subject-specific FNC for both fMRI and structural FNC.
We also tested CLiP-ICA on 864 fMRI and sMRI scans from ADNI. The fMRI data were reduced to 30 PCA (45 at the subject level and 30 at the group level), and the gray matter images were also reduced to 30 PCA. Then, we ran CLiP-ICA with a Gaussian copula, where linkage parameters ranged linearly from 0.95 to 0.65. After training, we compared the sMRI and fMRI components. Figure 2 shows the aggregated structural FNC results. Subsequently, utilizing back reconstruction enabled the calculation of subject-specific FNC for both fMRI and structural FNC.
·A component pair: on the left, an fMRI component; on the right, the corresponding linked sMRI component.
·Aggregated structural FNC between each fMRI and sMRI components
Conclusions:
In conclusion, this study introduces a groundbreaking algorithm, designed to fully leverage the complete 4D fMRI dataset in the fusion of fMRI and sMRI, thereby enabling enhanced linked independent component analysis. Our comprehensive simulations have demonstrated the superior performance of our proposed CLiP-ICA algorithm over the traditional Joint ICA (JICA) approach. Furthermore, the application of our method to real medical imaging datasets has validated its effectiveness in harnessing valuable information, marking a significant advancement in medical imaging analysis
Modeling and Analysis Methods:
Connectivity (eg. functional, effective, structural) 1
fMRI Connectivity and Network Modeling
Multivariate Approaches 2
Novel Imaging Acquisition Methods:
Anatomical MRI
BOLD fMRI
Keywords:
Degenerative Disease
FUNCTIONAL MRI
Machine Learning
Modeling
STRUCTURAL MRI
1|2Indicates the priority used for review