Inter-subject parcellation of functional Magnetic Resonance Imaging (fMRI) data based on a standard General Linear Model (GLM)and spectral clustering was recently proposed as a means to alleviate the issues associated with spatial normalization in fMRI. However, for all its appeal, a GLM-based parcellation approach introduces its own biases, in the form of a priori knowledge about the shape of Hemodynamic Response Function (HRF) and task-related signal changes, or about the subject behaviour during the task. In this paper, we introduce a data-driven version of the spectral clustering parcellation, based on Independent Component Analysis (ICA) and Partial Least Squares (PLS) instead of the GLM. First, a number of independent components are automatically selected. Seed voxels are then obtained from the associated ICA maps and we compute the PLS latent variables between the fMRI signal of the seed voxels (which covers regional variations of the HRF) and the principal components of the signal across all voxels. Finally, we parcellate all subjects data with a spectral clustering of the PLS latent variables. We present results of the application of the proposed method on both single-subject and multi-subject fMRI datasets. Preliminary experimental results, evaluated with intra-parcel variance of GLM t-values and PLS derived t-values, indicate that this data-driven approach offers improvement in terms of parcellation accuracy over GLM based techniques.
Deep Dive into Parcellation of fMRI Datasets with ICA and PLS-A Data Driven Approach.
Inter-subject parcellation of functional Magnetic Resonance Imaging (fMRI) data based on a standard General Linear Model (GLM)and spectral clustering was recently proposed as a means to alleviate the issues associated with spatial normalization in fMRI. However, for all its appeal, a GLM-based parcellation approach introduces its own biases, in the form of a priori knowledge about the shape of Hemodynamic Response Function (HRF) and task-related signal changes, or about the subject behaviour during the task. In this paper, we introduce a data-driven version of the spectral clustering parcellation, based on Independent Component Analysis (ICA) and Partial Least Squares (PLS) instead of the GLM. First, a number of independent components are automatically selected. Seed voxels are then obtained from the associated ICA maps and we compute the PLS latent variables between the fMRI signal of the seed voxels (which covers regional variations of the HRF) and the principal components of the sig
Inter-subject parcellation based on a standard General Linear Model (GLM) and spectral clustering was recently proposed as a means to alleviate the issues associated with spatial normalization in the analysis functional Magnetic Resonance Imaging (fMRI) datasets: lack of true anatomical correspondence, inaccuracy of the normalization process (see [1] for an in-depth overview), etc. In a parcellation framework, voxels are first clustered into functionally homogeneous regions or parcels. Then, the parcellations are then homogenised across subjects, so that statistics can be carried out at the parcel level rather than at the voxel level.
Here we focus on the optimization of the first step of the parcellation scheme.
We present a data-driven, model-free, parcellation technique, based on Independent Component Analysis (ICA) and Partial Least Squares (PLS, [2]) instead of a GLM, so as to use more of the information contained within the fMRI time series. First, a number of independent components are automatically selected. Seed voxels are then obtained from the associated ICA maps and we compute the PLS latent variables between the fMRI signal of the seed voxels (which covers regional variations of the stimuli related BOLD responses) and the principal components of the signal across all voxels. Finally, we parcellate all subjects data with a spectral clustering of the PLS latent variables. We also introduce PLS t-values as an alternative way to validate parcellation results.
We detail our approach in the following Section 2. Preliminary results are given in Section 3, where we also compare them to GLM-based parcellation.
We applied our method to two functional datasets: a single-subject fMRI experiment with a standard finger tapping task and a multi-subject experiment where volunteers were presented with hand gestures or face expressions [3].
Single-subject data were acquired on a Philips Intera 1.5T with a TR of 3s and a sequential finger tapping task auditorily paced with a metronome. The auditory signals were given every 0.6 seconds. The digit order of the tapping was 1 -3 -2 -4, repeated 6 times in each period, with a 14.4 second rest between periods. The period of one on-and-off block was then 28.8 seconds.
Multi-subject data, our main concern in this paper, were acquired from 25 subjects viewing angry gestures or expressions. Scanning was performed on a Philips Intera 1.5T, with TR=3s. During the scan, four types of visual stimuli are given to the subjects, which are angry hand gestures, neural hand gestures, angry facial expression and neural facial expression.
Both datasets were preprocessed with FSL for slice-timing, motion correction and registration [4,5].
For the multi-subject experiment, we used FSL to decompose the input fMRI data into independent components (ICs). We obtained between 30 and 60 ICs per subject, for a total of 1203 ICs. Here we propose to use a hierarchical clustering approach, similar in spirit to Partner Matching [6] as a means to find the ICs that best capture the Blood Oxygen Level Dependent Haemodynamic (BOLD) response to the stimuli. This method is based on the assumption that very few of the 1203 ICs contain information about the stimuli-related BOLD responses. Consequently, the task-related ICs should be more similar with each other than with the other ICs since they share the same source. We aim to group those ICs that correspond to the response to the same task features in different subjects = q A i i 3 together into one cluster. The other ICs which do not contain relevant (i.e. taskrelated) information should be grouped inside another cluster. We take those constraints into account when design the similarity function to be used in our hierarchical clustering.
Let N a and N b be the number of ICs for subjects A and B respectively, with IC A and IC B the ith IC of subject A and jth IC of subject B. t = 1, . . . , T is i j the time index. Their correlation coefficients is given by:
The normalized correlation coefficients ρ norm is:
Since the aim of the clustering is to put similar ICs from different subjects into one cluster, all the ICs of the same cluster should come from different subjects, therefore we need to set the similarity between ICs of the same subject to 0. The similarity between two ICs is finally defined as
(3) i , IC j ), ρ norm (IC j , IC i )) other wise .
In the case of the single-subject data, the ICs representing BOLD signals could not be selected by comparing ICs across subjects as above. Therefore, we manually picked those ICs that best matches the canonical HRF-convoluted task design from the 34 ICs produced by FSL.
In order to calculate the PLS latent variables that best capture the BOLD response, a number of seeds representing different active regions should be selected. For instance, in a GLM-based parcellation approach, we could select as seeds the voxels with the largest t-values. Here, we pick them on the basis of the ICA results.
Wi
…(Full text truncated)…
This content is AI-processed based on ArXiv data.
