Harmonization of multi-site MRI data with dual-projection based Linked ICA model
Poster No:
1873
Submission Type:
Abstract Submission
Authors:
Huashuai Xu1, Yuxing Hao1, Huanjie Li2, Fengyu Cong2, Tommi Kärkkäinen1
Institutions:
1Faculty of Information Technology, University of Jyväskylä, Jyväskylä, Finland, 2School of Biomedical Engineering, Faculty of Medicine, Dalian University of Technology, Dalian, China
First Author:
Co-Author(s):
Huanjie Li
School of Biomedical Engineering, Faculty of Medicine, Dalian University of Technology
Dalian, China
School of Biomedical Engineering, Faculty of Medicine, Dalian University of Technology
Dalian, China
Fengyu Cong
School of Biomedical Engineering, Faculty of Medicine, Dalian University of Technology
Dalian, China
School of Biomedical Engineering, Faculty of Medicine, Dalian University of Technology
Dalian, China
Introduction:
Merging magnetic resonance imaging (MRI) data from multiple sites has become increasingly popular. However, measurement biases caused by site differences in scanners represent a barrier while pooling data collected from different sites. The existence of site effects can mask biological effects and lead to false findings [2][4][5]. We proposed a dual-projection (DP) method based on the independent component analysis (ICA) method in our previous study [2]. The method demonstrated superior performance in mitigating site effects while preserving biological signals in the context of unimodal MRI data. This study proposes a novel multimodal denoising approach that implements a linked LICA -based DP method to remove the site effects. This method can separate the signal effects from the identified site effects and remove the site effects more completely. A dataset from Autism Brain Imaging Data Exchange II was used to test the proposed LICA-DP denoising method. The results indicate that the LICA-based DP method can remove site effects while preserving true biological variability.
Methods:
Data and preprocess
We utilized two modalities of functional MRI data, including amplitude of low frequency fluctuation (ALFF), and regional homogeneity (ReHo). The data were from Autism Brain Imaging Data Exchange II (ABIDE II). There were 795 subjects (including Autism Spectrum Disorder (ASD) patients: 341, Healthy Controls (HC): 454).
The raw fMRI data were preprocessed with FSL FEAT (https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/FEAT ), including removing the first six volumes, motion correction, and spatial normalization to standard MNI space. Two functional modalities, ALFF and ReHo, were generated from the preprocessed fMRI data with DPABI [6]. For ReHo, spatial smoothing (with Full Width at Half Maximum (FWHM) of 6 mm) was performed after ReHo calculation, but for ALFF, spatial smoothing was completed before the calculation [3].
LICA-DP method
The LICA [1] toolbox can be downloaded from the FSL website (https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/FLICA ). To preserve the signal effects, the traditional LICA method (LICA-SP) only removes those pure site-related components (related to site effects only), and leaves those mixed components without any process. To eradicate the site effects, we propose the LICA-DP method. Firstly, LICA-DP separates the signal effects out from the mixed components. The left noise contributions of the mixed components and the pure site-related components are regarded as the total site effects, which are then cleaned from the non-denoised data via a second projection procedure.
We utilized two modalities of functional MRI data, including amplitude of low frequency fluctuation (ALFF), and regional homogeneity (ReHo). The data were from Autism Brain Imaging Data Exchange II (ABIDE II). There were 795 subjects (including Autism Spectrum Disorder (ASD) patients: 341, Healthy Controls (HC): 454).
The raw fMRI data were preprocessed with FSL FEAT (https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/FEAT ), including removing the first six volumes, motion correction, and spatial normalization to standard MNI space. Two functional modalities, ALFF and ReHo, were generated from the preprocessed fMRI data with DPABI [6]. For ReHo, spatial smoothing (with Full Width at Half Maximum (FWHM) of 6 mm) was performed after ReHo calculation, but for ALFF, spatial smoothing was completed before the calculation [3].
LICA-DP method
The LICA [1] toolbox can be downloaded from the FSL website (https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/FLICA ). To preserve the signal effects, the traditional LICA method (LICA-SP) only removes those pure site-related components (related to site effects only), and leaves those mixed components without any process. To eradicate the site effects, we propose the LICA-DP method. Firstly, LICA-DP separates the signal effects out from the mixed components. The left noise contributions of the mixed components and the pure site-related components are regarded as the total site effects, which are then cleaned from the non-denoised data via a second projection procedure.
Results:
Figure 1 shows an evaluative exploration of the efficacy of different modalities to attenuate site effects. Compared to a unimodal approach, integrating ALFF and ReHo modalities can more effectively eliminate site effects. The LICA-SP method partially mitigates site effects (ALFF+ReHo > unimodal), whereas the LICA-DP approach can eliminate these effects (not impacted by the number of modalities involved).
Figure 2 presents the group-level analyses of age effects. The original data only allowed for detecting positive age effects, while negative age effects went undetected. This limitation can be attributed to site effects, which appear to hinder the detection of the negative age-related influence. After harmonization, notably by LICA-DP, we unveiled negative age effects that were not discernible in the non-denoised data.
Figure 2 presents the group-level analyses of age effects. The original data only allowed for detecting positive age effects, while negative age effects went undetected. This limitation can be attributed to site effects, which appear to hinder the detection of the negative age-related influence. After harmonization, notably by LICA-DP, we unveiled negative age effects that were not discernible in the non-denoised data.
Conclusions:
To tackle the shortcomings of not being able to eliminate site effects from the traditional LICA method, we propose a dual-projection data-driven method based on LICA. Compared to traditional LICA methods, which cannot thoroughly eliminate site effects while preserving biological signals, our LICA-DP approach effectively removes site effects and retains biological signals in single and multimodal data scenarios. On the other hand, multimodal fusion yields superior(or at least equivalent) multi-site harmonization results compared to unimodal data.
Modeling and Analysis Methods:
Methods Development 1
Multivariate Approaches 2
Keywords:
MRI
1|2Indicates the priority used for review
Provide references using author date format
2. Hao, Y. et al. (2023) ‘Removal of site effects and enhancement of signal using dual projection independent component analysis for pooling multi‐site MRI data’, European Journal of Neuroscience, 58(6), pp. 3466–3487. Available at: https://doi.org/10.1111/ejn.16120.
3. Jia, X.Z. et al. (2019) ‘RESTplus: an improved toolkit for resting-state functional magnetic resonance imaging data processing’, Science Bulletin, 64(14), pp. 953–954. Available at: https://doi.org/10.1016/j.scib.2019.05.008.
4. Li, H. et al. (2020) ‘Denoising scanner effects from multimodal MRI data using linked independent component analysis’, NeuroImage, 208(116388). Available at: https://doi.org/10.1016/j.neuroimage.2019.116388.
5. Xu, H. et al. (2023) ‘Harmonization of multi-site functional MRI data with dual-projection based ICA model’, Frontiers in Neuroscience, 17. Available at: https://doi.org/10.3389/fnins.2023.1225606.
6. Yan, C.G. et al. (2016) ‘DPABI: Data Processing & Analysis for (Resting-State) Brain Imaging’, Neuroinformatics, 14(3), pp. 339–351. Available at: https://doi.org/10.1007/s12021-016-9299-4