COMPUTATIONAL MEANS, CURRENT RESEARCH, AND TOPICS OF INTEREST

Radu Mutihac

Operating systems:

Programming languages:

Data type and analysis methods:

Software packages dedicated to biomedical time-series processing currently using:

• (Functional MRI Analysis Software Packages)
  1. SPM 2 & SPM 5 - Statistical Parametric Mapping -
     SPM 2 plus Wavelets analysis (WaveLab 802 designed by D. Donoho at Stanford University)
     SPM 2 plus SnPM - Statistical non-Parametric Mapping -
     SPM 2 plus WSPM (Wavelet-based SPM designed by D. Van de Ville, Biomedical Image Group, EPFL, Swiss)
  2. SourceForge EEG/MRI Toolbox for MATLAB;
  3. FSL 3.3 - Library of functional and structural brain image analysis tools - including FEAT (hypothesis-driven fMRI data analysis) and MELODIC (data-driven probabilistic ICA of fMRI) developed by Image Analysis Group, FMRIB, Oxford, UK.
  4. Evident 6.1 - 3D Functional Imaging Analysis Software Package running EROICA - exploring regions of interest with cluster analysis in large fMRI data sets - commercial product from Institute for Biodiagnostics, National Research Council of Canada;
  5. BrainVoyager 2000 v.4.9 - highly optimized software package for the analysis and visualization of functional and structural MRI data sets - commercial product
  6. GIFT 1.3b - Group ICA of fMRI toolbox - developed at the Institute of Living/Hartford Hospital and Yale University School of Medicine by Dr. Vince Calhoun;
  7. FMRLAB 3.0 - MATLAB toolbox for fMRI data analysis using extended infomax ICA - developed initially at the Computational Neurobiology Lab of The Salk Institute for Biological studies, now centered at the Swartz Center for Computational Neuroscience of the Institute for Neural Computation of the University of California San Diego.
  8. LYNGBY 1.05 (January 21, 2004) developed at IMM - Informatics and Mathematical Modelling DTU - Technical University of Denmark.
  9. ICALAB 2.2 - MATLAB toolboxes for signal pre-processing and post-processing developed at the RIKEN Laboratory for Advanced Brain Signal Processing, Japan.
• (Functional image visualizing tools)
  1. AFNI 2.56d (June 06, 2004) - Analysis of Functional NeuroImages designed at Medical College of Wisconsin - under CygWin, LINUX, and UNIX only.
  2. MEDx 3.4.3 - Image Processing Software Package - from Medical Numerics, Inc., under UNIX only.
  3. MRIcro 1.31 for visualization and converting both 3D and 4D functional neuroimages designed at Biomedical Imaging Resource, Mayo Clinic, WinXP only.

Current research and topics of interest in fMRI time-series processing:

1. Exploratory analysis of both synthetic and real fMRI data, mainly spatial ICA (sICA) and fuzzy clustering analysis (FCA) that were quantitatively compared, in search for an optimal sequence of data-driven model-free methods and their subsequent statistical significance assessment by hypothesis-driven procedures. Since cluster analysis operates on statistically selected active time courses of activation while sICA separates the independent components on the basis of their spatial independence, validation of ICA may consist in getting similar results by these theoretically distinct approaches. In brain imaging the preliminary results have indicated that no statistically significant difference exists in the activation areas revealed by both FCA and ICA, as far as the contrast-to-noise ratio (CNR) of fMRI signal is beyond a certain value. This study requires the commercial package EVIDENT running EROICA for exploring ROIs with (Bezdek's) fuzzy cluster analysis in three main stages. Statistical validation of the results obtained by FCA of fMRI data is to be performed by a permutation-based resampling technique, which allows for statistical significance assessment of each voxel belonging to a cluster of interest without parametric distributional assumption. In this framework, the null hypothesis corresponds to "no activation", while the alternative hypothesis is represented by a relevant cluster centroid from FCA. Pearson's correlation coefficient between the cluster centroid and each TC belonging to the "activation" cluster is to be chosen as test statistic. The distribution of the correlation coefficient is given by permutation of the time point labels of the cluster centroid. The generated distribution will be used to assess the statistical validity at a specified level of significance.



2. Systematic study on spatiotemporal ICA of fMRI data. ICA may be used in two complementary ways to decompose an image sequence into a set of images and a corresponding set of time-varying image amplitudes: (i) spatial ICA (sICA), which finds a set of mutually independent images and a corresponding dual set of unconstrained time courses, and (ii) temporal ICA (tICA), which finds a set of independent time courses and a corresponding (dual) set of unconstrained images. A reasonable approach to ICA is to enforce both spatial and temporal constraints such as maximizing the degree of independence of dual signals. Spatiotemporal ICA (stICA) maximizes the degree of independence over space and time simultaneously, without necessarily producing independence in either space or time, rather searching for the best compromise. So stICA yields solutions in which the degree of spatial independence is maximized subject to the constraint that the degree of temporal independence is maximized (and reciprocally). It turns out that stICA is based on the assumption that both spatial and temporal sources are almost independent. Such a relaxed assumption over the approximate independence of both the estimated components and their dual signals permits the recovery of sources that are correlated over time and space, which happens to be the case in functional neuroimaging.



3. Model selection of fMRI data. It refers to determining the size as well as the complexity of the data model subspace. True dimensionality of fMRI data subspace is conceptualized as the number of independent activations and non-Gaussian noise sources in the presence of Gaussian noise. Underestimation of the dimensionality discards valuable information resulting in suboptimal signal extraction. Overfitting a noise-free generative model to noisy observations (e.g., square ICA) results in a large number of spurious components due to unconstrained estimation and factorization, negatively affecting subsequent statistical inference, physiological interpretation, and increasing the computational demand. In all cases, if we relax the constraint on the ICA model of being square, then a mismatch between the best linear model fit and the original data is inevitably introduced. The envisaged approach is based on the theoretic-information concept of mutual information (MI) applied to a full square spatial ICA decomposition of fMRI data that were minimally preprocessed (slice timing correction, realign, and coregistration - in this order). If the mean MI plot displays a flex point, then it estimates the number of distinct spatial clusters which equals the number of distinct (latent) source signals. The results can easily be compared with some other multivariate methods of estimating the information/structure content like minimum description length (MDL), Akaike's information criterion (AIC), or probabilistic principal component analysis (pPCA).



4. Statistical assessment of ICA separated components. Noise in fMRI data is spatially correlated, but noise and data are assumed to be uncorrelated. In square ICA, the estimated ICs are determined solely by the data and the estimation of the mixing matrix. This precludes: (i) the assessment of statistical significance of ICA estimates in testing against a null-hypothesis, (ii) likewise, the threshold techniques like converting the component map values to z-scores is devoid of statistical meaning and can only be conceived as ad-hoc recipes. Since a noise model is missing, any slight differences in the measured hemodynamic response at two different voxel locations must be treated as real effects. However, these differences may represent either real spatial variations, or differences in the background noise level. If so, clusters of voxels activated by the same external stimulus may be split onto different spatial ICA maps raising the question of analysis validity, and subsequently increasing the difficulty of neurophysiologic interpretation and the overall computational cost.



5. A comparative study of Gaussian spatial smoothing versus wavelet denoising in fMRI data pre-processing, prior to spatial normalization and segmentation. Though most statistical tests applied to fMRI data assume Gaussian distributed noise under the null hypothesis, evidence exists that noise in MR images is Rician distributed. Rician noise is multiplicative in contrast with the Gaussian noise which is additive. However, analyzing the BOLD contrast as the difference between two MR images (e.g., active-baseline), both containing Rician distributed noise, the distribution of BOLD noise was found to closely approximate a Gaussian. In terms of data smoothing, SPM basically applies a spatial Gaussian filter to the data before performing statistical analysis. It degrades the image resolution and complicates the statistical analysis since the noise can no longer be considered independent. Familywise error (FEW) is the standard measure of type I errors in multiple hypothesis testing, which specifies the chance of false positives. SPM2 is endowed also with False Detection Rate (FDR), which is a metric for measuring the expected proportion of false positives (i.e., type I errors) among suprathreshold voxels (i.e., rejected null hypotheses) and which does not require spatial smoothness. Then wavelet denoising can be used with manifold advantages: (i) the compression effect tends to pack the signal into a relatively few number of large coefficients. The noise is evenly distributed in the wavelet domain, so that the signal-to-noise ratio (SNR) is improved for those coefficients where the signal concentrated; (ii) higher SNR improves the detection rate; and (iii) the wavelet transform is a representation with no redundancy; the decoupling effect in the wavelet domain allows a better control of the false-detection rate and ensures validity of the method even when the hypothesis of noise independence breaks down.



6. Alphanumeric brain patterns. A preliminary study of brain activity patterns for subjects visually trained with single normalized alpha-numeric characters ending up with an "alphabet" of brain elementary patterns. Then an experimental paradigm consisting of simple written words and/or short multidigit numbers may explain the generation of more complex brain patterns by superposition of elementary ones grouped in this "alphabet."



7. Tensorial ICA. The goal is extracting relevant information based on different imaging modalities from fMRI data simultaneously acquired from one subject, such as originating from contrasts like BOLD (Blood Oxygenation Level Dependent), VASO (Vascular Space Occupancy), and ASL (Arterial Spin Labeling) perfusion. A tensorial ICA model of fMRI data giving a concurrent decomposition into modes of variation across the spatial, temporal, and imaging modality domain in the presence of Gaussian noise may be of interest in simultaneous acquisition of complementary functional hemodynamic indices reflecting different aspects of brain activity and allowing enhanced detection power and improved data interpretation. Up to date, a novel MRI technique was reported that is claimed to achieve concurrent acquisition of three hemodynamic images based primarily on the changes of CBV (Cerebral Blood Volume), blood flow (CBF), and blood oxygenation (BOLD). Functional images based on different contrasts are usually acquired separately, though ASL, perfusion, and BOLD signals can be acquired in a single scan. The feasibility and efficacy of simultaneously recording of three imaging contrasts were assessed by brain activation experiments with visual stimulation paradigms; higher contrast-to-noise ratio per unit time compared with conventional techniques collecting these functional images separately was reported. However, data analysis was classically carried out, that is, separately for each imaging modality, which is computationally demanding and time consuming, at least. A 3D approach to ICA (e.g., space, time, imaging modality) resorting to tensorial algebra would computationally complement the advent of simultaneous acquisition of fMRI data based on three different hemodynamic images with many advantages concerning: (i) efficient acquisition of multiple functional images, (ii) minimization of temporal fluctuations in the time series of these images, (iii) a combination of complementary functional images may offer a better understanding of the physiological and/or biophysical transduction mechanisms among neural activity, hemodynamics, and MRI signals, (iv) CBF, CBV, and BOLD images indicate the cerebral metabolic rate of oxygen and, implicitly, the regional oxygen consumption and metabolism, a measure potentially less sensitive to vascular dynamics.



8. Coordinate converter from SPM-MNI space to stereotactic space and automated anatomic labeling. Many statistical software packages for fMRI data analysis output functional images in the SPM-MNI space. Anatomic labeling of various regions of interest (ROIs) is feasible using Talairach Daemon (TD) database within the WFU PickAtlas software toolbox (developed in the Functional MRI Laboratory at the Wake Forest University School of Medicine), which provides a method for generating ROI masks. The segmented atlases in WFU PickAtlas are in the MNI_atlas_templates subdirectory with their corresponding look-up tables. The atlases are in MNI space with dimensions of 91x109x91 sampled at 2 mm intervals, corresponding to the SPM MNI templates. The atlases are in neurologic convention (right of image = right of subject). The other way around is to plug in the output ROIs yielded by SPM of fMRI data to TD and get the anatomical labels. In order to perform automated anatomic labeling, a nonlinear transform from SPM-MNI coordinates to Talairach space is necessary. An attempt in this respect has already been described by Matthew Brett (http://www.mrc-cbu.cam.ac.uk/Imaging/Common/mnispace.shtml), yet no validation of this transform has been reported so far.



9. Wavelet-based statistical analysis of neuroimaging data. Wavelet analysis is optimal in terms of detecting transients events in fMRI time series and adapts well to conditions where responses change significantly in amplitude during experiments. Wavelet-based methods provide a naturally multiscale alternative to single scale Gaussian spatial smoothing as widely used before hypothesis testing. Scale-varying wavelet-based methods for hypothesis testing of brain activation maps circumvent the need to specify a priori the size of signals expected and, therefore, the optimal choice of the smoothing kernel required by Gaussian filtering. Due to the smoothness of the wavelet representation, the estimated statistical parameter maps reveal more compact regions of activation than their counterparts obtained by statistic testing in the spatial domain. Wavelet-based methods are likely to provide an overall richer characterization of distributed brain activation. A comparison of various wavelet bases following the criterion of their efficiency in terms of the best peak-to-signal-to-noise ratio (PSNR) in image denoising applications is intended. So far, our results have indicated that the orthogonal fractional B-splines are a near-optimal choice of the wavelet bases and that their parameters can be optimized by applying the SURE principle. Moreover, the addition of a complex part to these parameters significantly improves the results, which is hardly the case for the complex extension of other commonly used wavelets.

Last updated:
Bucharest, May 30, 2006