Robust individual analysis of multi-band multi-echo functional MRI
Presenting author:
To perform functional MRI (fMRI) rapidly one can utilise multi-band (simultaneous multi-slice) fMRI. To improve signal to noise, one can acquire each image at several echo times (TE). One can further take advantage of multiple echo data, exploiting echo-time dependency of the blood-oxygenation level dependent fMRI signal of interest; TE Dependent ANAlysis (TEDANA; https://tedana.readthedocs.io) is an existing workflow to achieve this, using independent component analysis (ICA), together with echo-dependent signal classification, to denoise the fMRI data. We wished to determine whether this automated approach is sufficiently robust to reliably yield acceptable results for individual subjects, a necessary pre-requisite for our clinical research.
We tested this approach in a pilot study for the Australian Epilepsy Project (AEP). Multi-band multi-echo fMRI was acquired, with a sequence from the University of Minnesota CMRR, in 196 participants performing a language task. The data were initially pre-processed using fMRIPrep (https://fmriprep.org). Statistical analysis used the iBrain Analysis Toolbox for SPM (https://florey.edu.au/iBrain) and SPM12 (https://www.fil.ion.ucl.ac.uk/spm). To quantify TEDANA performance, we calculated mean t-score and activation volume within a language region of interest. Our analysis showed that, for a handful of subjects, results were unsatisfactory: TEDANA removed not just noise, but also considerable neural activity. We hypothesised that low performance is caused by 1) inadequate thermal noise suppression, 2) dependency of FastICA (used by TEDANA) on an initial seed value (the result could change substantially depending on the seed) and 3) misclassification, as noise, of some ICA components with substantial neuronal signal.
We modified the mutli-echo denoising pipeline to address all these issues: 1) applied Marchenko-Pastur PCA (MP-PCA) to fMRI raw data (prior to fMRIPrep) and bypassed the then redundant moving-average PCA dimensionality-reduction step in the TEDANA pipeline, 2) replaced FastICA with RobustICA (https://github.com/CRG-CNAG/robustica) that provides reliable ICA components by re-calculating over many initial seed values and 3) modified the classification algorithm.
Maximum performance was achieved only with all these modifications in-place, at a group as well as individual subject level. The improved performance of the new method was confirmed by clinicians’ evaluation, and on independent data.
We tested this approach in a pilot study for the Australian Epilepsy Project (AEP). Multi-band multi-echo fMRI was acquired, with a sequence from the University of Minnesota CMRR, in 196 participants performing a language task. The data were initially pre-processed using fMRIPrep (https://fmriprep.org). Statistical analysis used the iBrain Analysis Toolbox for SPM (https://florey.edu.au/iBrain) and SPM12 (https://www.fil.ion.ucl.ac.uk/spm). To quantify TEDANA performance, we calculated mean t-score and activation volume within a language region of interest. Our analysis showed that, for a handful of subjects, results were unsatisfactory: TEDANA removed not just noise, but also considerable neural activity. We hypothesised that low performance is caused by 1) inadequate thermal noise suppression, 2) dependency of FastICA (used by TEDANA) on an initial seed value (the result could change substantially depending on the seed) and 3) misclassification, as noise, of some ICA components with substantial neuronal signal.
We modified the mutli-echo denoising pipeline to address all these issues: 1) applied Marchenko-Pastur PCA (MP-PCA) to fMRI raw data (prior to fMRIPrep) and bypassed the then redundant moving-average PCA dimensionality-reduction step in the TEDANA pipeline, 2) replaced FastICA with RobustICA (https://github.com/CRG-CNAG/robustica) that provides reliable ICA components by re-calculating over many initial seed values and 3) modified the classification algorithm.
Maximum performance was achieved only with all these modifications in-place, at a group as well as individual subject level. The improved performance of the new method was confirmed by clinicians’ evaluation, and on independent data.