Haut PDF Subject-level Joint Parcellation-Detection-Estimation in fMRI

Subject-level Joint Parcellation-Detection-Estimation in fMRI

Subject-level Joint Parcellation-Detection-Estimation in fMRI

Subject-level Joint Parcellation-Detection-Estimation in fMRI Lotfi CHAARI Member, Solveig BADILLO, Thomas VINCENT, Ghislaine DEHAENE-LAMBERTZ, Florence FORBES and Philippe CIUCIU Senior Member Abstract—Brain parcellation is one of the most important issues in functional MRI (fMRI) data analysis. This parcellation allows establish- ing homogeneous territories that share the same functional properties. This paper presents a model-based approach to perform a subject-level parcellation into hemodynamic territories with similar hemodynamic features which are known to vary between brain regions. We specifically investigate the use of the Joint Parcellation-Detection-Estimation (JPDE) model initially proposed in [1] to separate brain regions that match different hemodynamic response function (HRF) profiles. A hierarchi- cal Bayesian model is built and a variational expectation maximiza- tion (VEM) algorithm is deployed to perform inference. A more complete version of the JPDE model is detailed. Validation on synthetic data shows the robustness of this model to varying signal-to-noise ratio (SNR) as well as to different initializations. Our results also demonstrate that good parcellation performance is achieved even though the parcels do not involve the same amount of activation. On real fMRI data acquired in children during a language paradigm, we retrieved a parcellation along the superior temporal sulcus of the left hemisphere that matches the gradient of activation dynamics already reported in the literature.
En savoir plus

11 En savoir plus

Hemodynamic-Informed Parcellation of fMRI Data in a Joint Detection Estimation Framework

Hemodynamic-Informed Parcellation of fMRI Data in a Joint Detection Estimation Framework

that an accurate HRF model may significantly improve detection performance. To capture this variability, robust HRF estimation is necessary which can be achieved only in voxels or regions that elicit an evoked response to a given stimulus [2]. So far, many works have addressed this issue either by considering linear or nonlinear HRF models [3–5], parametric, semi-parametric or non-parametric (i.e. FIR models) descriptions [6–8], and by performing univariate (voxelwise) [4, 7], multivariate (regionwise) [9, 10] or even multiscale, i.e. spatially adaptive inference [11]. However, to the best of our knowledge, all these existing works assume the spatial support of the HRFs, either defined at the voxel or region-level, to be pre-specified. The proposed methodology takes place in the Joint Detection-Estimation (JDE) framework introduced in [9] and extended in [10, 12, 13] to account for spatial correlation between voxels. Standard JDE-based inference requires a pre-specified decomposition of the brain into functionally homogeneous parcels (groups of connected voxels) but with no guarantee of their optimality. These parcels should be small enough to guarantee the invariance of the HRF within each parcel, but large enough to contain reliable information for its inference [14]. Several attempts have been conducted to provide a robust parcellation such as in [15–19]. However, these approaches do not fully account for hemodynamics variability. Here, we introduce the concept of hemodynamic territory as a set of parcels which share a common HRF pattern. To determine such sets, we incorporate an additional layer in the JDE hierarchy, namely an adaptive parcel identification step based upon local hemodynamic properties. In this novel Joint Parcellation-Detection-Estimation (JPDE) model (Section II), for all the parcels of a given territory, HRFs are voxelwise but defined as local stochastic perturbations of the same HRF pattern. Then, hemodynamics estimation reduces to the identification of a limited number (say K) of such HRF patterns and parcel identification reformulates as a clustering problem where each voxel is assigned an HRF group among K. The HRF group assignment variables are governed by a hidden Markov model to enforce spatial correlation, i.e. favor group assignments that are spatially homogeneous. Finally, the overall scheme iteratively identifies hemodynamic territories as pairs of one HRF pattern and a set of parcels assigned to the corresponding HRF group.
En savoir plus

31 En savoir plus

Multi-subject joint parcellation detection estimation in functional MRI

Multi-subject joint parcellation detection estimation in functional MRI

subject-level [4] where the ”P” means that an additional layer complements the model to infer the brain parcellation. However, so far, the JPDE model has not been used for multi-subject fMRI analysis. Besides, the JDE model has been used in a multi-subject context [6] but the parcellation remained fixed a priori and identical across subjects. Another approach was proposed in [2] based on a semi-parametric framework with the general linear model. This approach assumes that for a fixed voxel under a given stimulus, the HRFs share the same unknown functional form across subjects but with different characteristics such as the time to peak, height and width. This common functional form was estimated using a nonparametric spline-smoothing method. In this paper, we introduce a joint intra and inter-subject fMRI data analysis model in the JPDE framework.The latter allows us to estimate a group-level parcellation as well as the group-level underlying HRF in contrast with the JPDE model. This group-level JPDE also recovers evoked activity in each individual. Moreover, the analysis is carried out for all the parcels of a region of interest (ROI) contrary to [6] where the analysis is done for a specific parcel at a time. The rest of the paper is organized as follows; Section II recalls the JPDE model introduced in [4]. Section III de- scribes our Multi-Subject Joint Parcellation Detection Esti- mation (MS-JPDE) extension. Experiments on multi-subject synthetic and real data are shown in Section IV. Conclusions are drawn in Section V.
En savoir plus

5 En savoir plus

Multi-subject joint parcellation detection estimation in functional MRI

Multi-subject joint parcellation detection estimation in functional MRI

subject-level [4] where the ”P” means that an additional layer complements the model to infer the brain parcellation. However, so far, the JPDE model has not been used for multi-subject fMRI analysis. Besides, the JDE model has been used in a multi-subject context [6] but the parcellation remained fixed a priori and identical across subjects. Another approach was proposed in [2] based on a semi-parametric framework with the general linear model. This approach assumes that for a fixed voxel under a given stimulus, the HRFs share the same unknown functional form across subjects but with different characteristics such as the time to peak, height and width. This common functional form was estimated using a nonparametric spline-smoothing method. In this paper, we introduce a joint intra and inter-subject fMRI data analysis model in the JPDE framework.The latter allows us to estimate a group-level parcellation as well as the group-level underlying HRF in contrast with the JPDE model. This group-level JPDE also recovers evoked activity in each individual. Moreover, the analysis is carried out for all the parcels of a region of interest (ROI) contrary to [6] where the analysis is done for a specific parcel at a time. The rest of the paper is organized as follows; Section II recalls the JPDE model introduced in [4]. Section III de- scribes our Multi-Subject Joint Parcellation Detection Esti- mation (MS-JPDE) extension. Experiments on multi-subject synthetic and real data are shown in Section IV. Conclusions are drawn in Section V.
En savoir plus

5 En savoir plus

Fast joint detection-estimation of evoked brain activity in event-related fMRI using a variational approach

Fast joint detection-estimation of evoked brain activity in event-related fMRI using a variational approach

I. I NTRODUCTION Functional Magnetic Resonance Imaging (fMRI) is a powerful tool to non-invasively study the relation- ship between a sensory or cognitive task and the ensuing evoked neural activity through the neurovascular coupling measured by the BOLD signal [1]. Since the 90’s, this neuroimaging modality has become widely used in brain mapping as well as in functional connectivity study in order to probe the specialization and integration processes in sensory, motor and cognitive brain regions [2–4]. In this work, we focus on the recovery of localization and dynamics of local evoked activity, thus on specialized cerebral processes. In this setting, the key issue is the modeling of the link between stimulation events and the induced BOLD effect throughout the brain. Physiological non-linear models [5–8] are the most specific approaches to properly describe this link but their computational cost and their identifiability issues limit their use to a restricted number of specific regions and to a few experimental conditions. In contrast, the common approach, being the focus of this paper, rather relies on linear systems which appear more robust and tractable [2, 9]. Here, the link between stimulation and BOLD effect is modelled through a convolutive system where each stimulus event induces a BOLD response, via the convolution of the binary stimulus sequence with the Hemodynamic Response Function (HRF 1 ). There are two goals for such BOLD analysis: the detection of where cerebral activity occurs and the estimation of its dynamics through the HRF identification. Commonly, the estimation part is ignored and the HRF is fixed to a canonical shape which has been derived from human primary visual area BOLD response [10, 11]. The detection task is performed by a General Linear Model (GLM), where stimulus-induced components are assumed to be known and only their relative weighting are to be recovered in the form of effect maps [2]. However, spatial intra-subject and between-subject variability of the HRF has been highlighted [12–14], in addition to potential timing fluctuations induced by the paradigm (e.g. variations in delay [15]). To take this variability into account, more flexibility can be injected in the GLM framework by adding more regressors. In a parametric setting, this amounts to adding a function basis, such as canonical HRF derivatives, a set of gamma or logistic functions [15, 16]. In a non-parametric setting, all HRF coefficients are explicitly encoded as a Finite Impulse Response (FIR) [17]. The major drawback of these GLM extensions is the multiplicity of regressors for a given condition, so that the detection task becomes more difficult to perform and that statistical power is decreased. Moreover, the more coefficients to recover, the more ill-posed the problem becomes. The alternative approaches that aim at keeping a single regressor per condition and add also a temporal regularization constraint to fix the ill-posedness
En savoir plus

40 En savoir plus

Multi-session extension of the joint-detection framework in fMRI

Multi-session extension of the joint-detection framework in fMRI

Index Terms— Brain activity, hemodynamics, JDE, fMRI, Bayesian inference, multisession 1. INTRODUCTION In the context of fMRI data analyses, the present paper is a contribu- tion to encoding methods. In such studies, two main concerns arise at the subject-level analysis: (i) a precise localization of evoked brain activity elicited by sensorimotor or cognitive tasks, and (ii) a robust estimation of the underlying hemodynamic response associated with these activations. Since these two steps are inherently linked, the Joint Detection-Estimation (JDE) approach [1, 2], has been proposed to face these issues in a coordinated formalism. This approach per- forms a multivariate inference for both detection and estimation. It makes use of a regional bilinear generative model of the BOLD re- sponse and constrains parameter estimation by physiological priors using temporal and spatial information in a Markovian model. The efficiency and usefulness of this approach has been validated at the group level in [3] considering single-session datasets.
En savoir plus

5 En savoir plus

Group-level impacts of within- and between-subject hemodynamic variability in fMRI

Group-level impacts of within- and between-subject hemodynamic variability in fMRI

In the present paper, both accurate activation detection and robust HRF estimation are addressed. Detecting evoked brain activity is usually performed in the General Lin- ear Model (GLM) context and postulates an invariant or canonical form for the HRF throughout the brain. However, several contributions have exhibited empirically de- rived HRFs that differ significantly from “the canonical model” [2, 36, 38, 22], and demonstrated that using such HRFs may improve the statistical sensitivity. Thus, GLM extensions have been developed to account for spatial fluctuations of the HRF shape by considering temporal and dispersion derivatives of the canonical model [17, 24]. In some cases, this approach lacks flexibility to take large deviations from canonical features into account. To overcome this limitation, nonparametric Finite Impulse Re- sponse (FIR) models have been proposed in the GLM framework. These models do not assume any functional form for the HRF and amounts to estimating a large number of parameters to identify its characteristics properties [19, 25]. This FIR-based modeling thus makes it possible to perform activation detection and HRF estimation. However, the achieved gain has a direct counterpart, namely a cost in terms of statistical sensitiv- ity since it induces a loss of degrees of freedom [34]. Moreover, FIR-based inference may lack robustness and often yields disrupted shapes that are difficult to interpret, due to over-fitting problems. Hence, to cope with this issue, temporal [21, 11, 35, 5] but also spatial regularization [20, 52, 10, 33, 41, 16, 27, 13, 6, 51] have been introduced in models to improve estimation accuracy.
En savoir plus

35 En savoir plus

Fast joint detection-estimation of evoked brain activity in event-related fmri using a variational approach

Fast joint detection-estimation of evoked brain activity in event-related fmri using a variational approach

Besides, results on HRF estimates are reported for the two JDE versions and compared to the canonical HRF, as well as maps of regularization factor estimates. Fig. 12 shows results for the VA contrast. High positive values are bilaterally recovered in the occipital region and the overall cluster localizations are consistent for both MCMC and VEM algorithms. The only difference lies in the temporal auditory regions, especially on the right side, where VEM yields rather more negative values than MCMC. Thus VEM seems more sensitive than MCMC. The results obtained by the classical GLM (see Fig. 12 [right]) are comparable to those of JDE in the occipital region with roughly the same level of re- covered activations. However, in the central region, we observe activations in the white matter that can be interpreted as false positives and that were not exhibited using the JDE formalism. The bottom part of Fig. 12 compares the estimated values of the regularization factors between VEM and MCMC algorithms for two experimental conditions involved in the VA contrast. Since these estimates are only relevant in parcels which are activated by at least one condition, a mask was applied to hide non-activated parcels. We used the following criterion to classify a parcel as activated: (and nonactivated otherwise). These maps of estimates show that VEM yields more contrasted values between the visual and auditory conditions. Table III provides the estimated values in the highlighted parcels of interest. The auditory condition does not elicit evoked activity and yields lower values in both parcels whereas the visual condition is associated with higher values. The latter comment holds for both algorithms but VEM provides much lower values than MCMC for the non-activated condition. For the activated condition, the situation is comparable, with and . This illustrates a noteworthy difference between VEM and MCMC. Probably due to the mean Þeld and variational approximations, the hidden Þeld may not have the same behaviour (different regularization effect) between the two algorithms. Still, this discrepancy is not visible on the NRL maps.
En savoir plus

18 En savoir plus

IMPACT OF THE JOINT DETECTION-ESTIMATION APPROACH ON RANDOM EFFECTS GROUP STUDIES IN FMRI

IMPACT OF THE JOINT DETECTION-ESTIMATION APPROACH ON RANDOM EFFECTS GROUP STUDIES IN FMRI

This paper is structured as follows. For the sake of self- consistency, the classical fMRI analysis framework is sum- marized in Section 2. The JDE approach is presented in Sec- tion 3. It relies on a prior parcellation of fMRI data, which derives from a clustering procedure that preserves connectiv- ity and functional homogeneity. Then, at the parcel level the JDE framework allows us to specify and estimate a specific BOLD model. Section 4 is devoted to group studies in fMRI where the principles of random effect analysis are reminded. In Section 5, results obtained at the group level using different subject-level inferences are compared on two salient contrasts of interest of a quick fMRI mapping experiment. A special at- tention is paid to the HRF variability in the motor and parietal regions. Conclusions are drawn in Section 6.
En savoir plus

6 En savoir plus

Application and validation of spatial mixture modelling for the joint detection-estimation of brain activity in fMRI.

Application and validation of spatial mixture modelling for the joint detection-estimation of brain activity in fMRI.

J. Idier, Ph.D., IRCCyN (CNRS), 1 rue de la Noë, BP 92101 44321 Nantes cedex 3, France jerome.idier@irccyn.ec-nantes.fr (activating or non-activating). The parameter controlling the strength of the spatial correlation is set by hand, as the smoothing level used when spatially filtering the data. The combination of these prior distributions with the likelihood allows us to derive the target posterior distribution using Bayes’ rule. We then resort to Gibbs sampling to draw realizations from this posterior law. The posterior mean (PM) estimates of the HRF, the Neural Response Levels (NRLs) and the corresponding labels are directly computed from the generated samples in a Markov Chain Monte Carlo (MCMC) procedure. In [2], this SMM approach was shown to give a significant gain in terms of sensitivity and specificity on artificial fMRI data, compared to the IMM counterpart. Since our primary interest was the comparison of SMM with IMM, our simulations followed the generative model and did not truly reflect a “realistic” BOLD data structure. Here, we assess the robustness against putative misspecifications from the assumed model and present first results on real fMRI data obtained during an event-related paradigm.
En savoir plus

6 En savoir plus

Robust Extrapolation Scheme for Fast Estimation of 3D Ising Field Partition Functions: Application to Within-Subject fMRI Data Analysis

Robust Extrapolation Scheme for Fast Estimation of 3D Ising Field Partition Functions: Application to Within-Subject fMRI Data Analysis

1 Introduction In fMRI, one usually resorts to spatial filtering to enhance the signal-to-noise ratio at the expense of a loss of spatial resolution. A more challenging approach works on the unsmoothed data by introducing some prior knowledge on the sought spatial structures through for instance local interaction models such as Markov Random Fields (MRFs). Discrete MRFs, which have been used in seg- mentation and clustering, typically involve a set of hyper-parameters: the smaller this number the less complex the patterns modelled by the corresponding MRF. For instance, a single temperature level controls the amount of spatial correlation in symmetric Ising fields. In the considered fMRI application [1], such Ising fields are hidden since the activation detection process is modelled a priori through a two-class Spatial Mixture Model (SMM). Moreover, their definition varies within a brain parcellation that segregates the 4D data into Γ functionally homogeneous
En savoir plus

9 En savoir plus

Adaptive Mean Shift Based Hemodynamic Brain Parcellation in fMRI

Adaptive Mean Shift Based Hemodynamic Brain Parcellation in fMRI

The detection of the evoked activity and the estimation of the dynamics have been mainly addressed as two separate tasks while each of them depends on the other. A precise localization of activations depends on a reliable HRF estimate, while a robust HRF shape is only achievable in brain regions eliciting task-related activity [11,12]. In this context, the joint detection estimation (JDE) model per- forms both tasks simultaneously [13–15]. In the JDE model, a single HRF shape is considered for a specific parcel (group of voxels). Although the JDE model jointly detects the evoked activity within the brain and estimates the HRF, it still requires a prior parcellation of the brain into functionally homogeneous regions. This challenge motivated the development of the joint parcellation de- tection estimation (JPDE) model [16,17] that performs online parcellation along with the detection and estimation tasks by setting voxels that share the same HRF pattern in the same HRF group (parcel). The JPDE model can be inferred using the VEM algorithm. However, this model still requires manual settings of the number of parcels. To overcome this issue, a model selection procedure was proposed in [18] to select the optimum number of parcels. This procedure depends mainly on free energy calculations where the model that maximizes the free energy is the best fit for the data. The limitation of this procedure arises from the fact that it needs to be run for each candidate model which can be time consuming especially if no prior information exists about the number of parcels. The standard JPDE model has been adopted in a Bayesian non-parametric ap- proach [19] by making use of the Dirichlet process mixtures model combined with a hidden Markov random field to automatically infer the number of parcels and their shapes simultaneously with the estimation and detection tasks. In this paper, a new approach is proposed to estimate the number of parcels from the fMRI BOLD signal. More precisely, we propose to embed the adaptive mean shift algorithm (which is a common clustering algorithm) within the variational inference framework associated with the JPDE model to estimate the parcels and their corresponding HRF profiles.
En savoir plus

13 En savoir plus

Hemodynamically informed parcellation of cerebral FMRI data

Hemodynamically informed parcellation of cerebral FMRI data

Index Terms— joint detection-estimation, hemodynam- ics, Gaussian mixtures, parcellation, brain 1. INTRODUCTION Functional MRI (fMRI) is an imaging technique that indi- rectly measures neural activity through the Blood-oxygen- level-dependent (BOLD) signal [1], which captures the vari- ation in blood oxygenation arising from an external stimula- tion. This variation also allows the estimation of the under- lying dynamics, namely the characterization of the so-called hemodynamic response function (HRF). The hemodynamic characteristics are likely to spatially vary, but can be consid- ered constant up to a certain spatial extent. Hence, it makes sense to estimate a single HRF shape for any given area of the brain. To this end, parcel-based approaches that segment fMRI data into functionally homogeneous regions and per- form parcelwise fMRI data analysis provide an appealing framework [2].
En savoir plus

6 En savoir plus

A fully Bayesian approach to the parcel-based detection-estimation of brain activity in fMRI.

A fully Bayesian approach to the parcel-based detection-estimation of brain activity in fMRI.

From a methodological point of view, we have shown that our joint detection- estimation technique is able to identify deactivations in the brain. This is owing to the introduction of a third class in the prior mixture model associ- ated to the NRLs. Nonetheless, we did not exhibit real deactivations on the analysed datasets. In the future, we should therefore validate the 3-class exten- sion on specific datasets. A good candidate could be a dataset acquired during an event-related auditory paradigm in which silence events are presented ran- domly to compare activations to a baseline derived from such events. As al- ready shown in (Ciuciu et al., 2003), silence events may generate deactivations in the temporal lobe if they are presented to the subject when the gradients of the scanner are switched off. This will be the subject of further work. Smoothing the data spatially provides a reliable manner for recovering clusters of activation instead of isolated spots, at the expense of a loss of resolution. To avoid this preprocessing, the proposed method could be extended by in- troducing spatial correlation in the prior model. This could be done either on the NRLs ( a) or on the underlying states (labels q). We argue in favour of the second solution for simplicity reasons. As already derived for Gaussian mixtures in (Vincent et al., 2007b,a), it is quite simple to sample from an Ising (2-class model) or Potts (3-class model) Markov random field (MRF) that enforce neighbouring voxels to be classified in the same state (e.g., ac- tivating). This approach actually seems more reasonable in terms of compu- tational load than considering edge-preserving MRF based on non-quadratic potentials (Green, 1990; Geman and McClure, 1987). Also, for computational reasons this extension has been developed in a supervised framework meaning that the hyper-parameter encoding spatial regularity of the hidden MRF is set by hand. Future work will be focused on an spatially adaptive extension in which this parameter is estimated as well, as already done in (Woolrich et al., 2005; Woolrich and Behrens, 2006).
En savoir plus

60 En savoir plus

A group model for stable multi-subject ICA on fMRI datasets.

A group model for stable multi-subject ICA on fMRI datasets.

In this paper, we present a novel group model for multivariate patterns in fMRI volumes and an associated estimation procedure to extract group-level ICA maps modeling subject variability. The strength of this method called CanICA lies in the identification of a subspace of reproducible components across subjects using generalized canonical correlation analysis (CCA). Combined with an explicit noise model and resampling procedure, this enables automatic selection of the number of components. In addition, we introduce a cross-validation procedure and metrics to compare the stability of a set of multi-subject patterns across different sub-populations. We compare our method to state-of-the-art fMRI group ICA methods with different group models: concatenation and tensorial group ICA approaches. We do not compare to merging procedures since they do not rely on a linear model between individual subject-level datasets and group-level Independent Components (ICs) and thus cannot be formulated with a spatially- resolved between-subject variability of group-level ICs. We show with cross-validation that features extracted by our method are more stable on a group of 12 controls, both in a resting-state experiment and in a traditional activation detection experiment with a known paradigm.
En savoir plus

29 En savoir plus

Impact of the joint detection-estimation approach on random effects group studies in fMRI

Impact of the joint detection-estimation approach on random effects group studies in fMRI

This paper is structured as follows. For the sake of self- consistency, the classical fMRI analysis framework is sum- marized in Section 2. The JDE approach is presented in Sec- tion ??. It relies on a prior parcellation of fMRI data, which derives from a clustering procedure that preserves connectiv- ity and functional homogeneity. Then, at the parcel level the JDE framework allows us to specify and estimate a specific BOLD model. Section ?? is devoted to group studies in fMRI where the principles of random effect analysis are reminded. In Section ??, results obtained at the group level using dif- ferent subject-level inferences are compared on two salient contrasts of interest of a quick fMRI mapping experiment. A special attention is paid to the HRF variability in the motor and parietal regions. Conclusions are drawn in Section ??.
En savoir plus

6 En savoir plus

Bayesian Joint Detection-Estimation of cerebral vasoreactivity from ASL fMRI data

Bayesian Joint Detection-Estimation of cerebral vasoreactivity from ASL fMRI data

BM NeuroSpin center, Bˆ at. 145, F-91191 Gif-sur-Yvette, France Abstract. Although the study of cerebral vasoreactivity using fMRI is mainly conducted through the BOLD fMRI modality, owing to its relatively high signal-to-noise ratio (SNR), ASL fMRI provides a more interpretable measure of cerebral vasoreactivity than BOLD fMRI. Still, ASL suffers from a low SNR and is hampered by a large amount of physiological noise. The current contribution aims at improving the re- covery of the vasoreactive component from the ASL signal. To this end, a Bayesian hierarchical model is proposed, enabling the recovery of per- fusion levels as well as fitting their dynamics. On a single-subject ASL real data set involving perfusion changes induced by hypercapnia, the approach is compared with a classical GLM-based analysis. A better goodness-of-fit is achieved, especially in the transitions between baseline and hypercapnia periods. Also, perfusion levels are recovered with higher sensitivity and show a better contrast between gray- and white matter.
En savoir plus

9 En savoir plus

A joint detection-estimation framework for analysing within-subject fMRI data

A joint detection-estimation framework for analysing within-subject fMRI data

Titre: Un cadre de détection-estimation conjointe pour analyser les données individuelles d’IRMf Philippe Ciuciu 1 , Thomas Vincent 1 , Laurent Risser 2 and Sophie Donnet 3 Abstract: In this paper, we review classical and advanced methodologies for analysing within-subject functional Magnetic Resonance Imaging (fMRI) data. Such data are usually acquired during sensory or cognitive experiments that aims at stimulating the subject in the scanner and eliciting evoked brain activity. From such four-dimensional datasets (three in space, one in time), the goal is twofold: first, detecting brain regions involved in the sensory or cognitives processes that the experimental design manipulates; second, estimating the underlying activation dynamics. The first issue is usually addressed in the context of the General Linear Model (GLM), which a priori assumes a functional form for the impulse response of the hemodynamic filter. The second question aims at estimating this shape which makes sense in activating regions only. In the last five years, a novel Joint Detection-Estimation (JDE) framework addressing these two questions simultaneously has been proposed in [59, 60, 102]. We show to which extent this methodology outperforms the GLM approach in terms of statistical sensitivity and specificity, which additional questions it allows us to investigate theoretically and how it provides us with a well-adapted framework to treat spatially unsmoothed real fMRI data both in the 3D acquisition volume and on the cortical surface.
En savoir plus

32 En savoir plus

Robust voxel-wise Joint Detection Estimation of Brain Activity in fMRI

Robust voxel-wise Joint Detection Estimation of Brain Activity in fMRI

Index Terms— Variational EM, MRF, Biomedical signal detection, Magnetic resonance imaging. 1. INTRODUCTION Functional Magnetic Resonance Imaging (fMRI) is a powerful tool to non-invasively study the relation between cognitive task and cere- bral activity through the analysis of the hemodynamic BOLD sig- nal [4]. Within-subject analysis in event-related fMRI first relies on (i) a detection step to localize which parts of the brain are acti- vated by a given stimulus type, and second on (ii) an estimation step to recover the temporal dynamics of the brain response. Most ap- proaches to detect neural activity rely on a single a priori model for the temporal dynamics of activated voxels also known as the hemodynamic response function (HRF) [5]. A canonical HRF is usually assumed for the whole brain although there has been evi- dence that this response can vary in space or region, across subjects and groups [6]. In addition, a robust and accurate estimation of the HRF is possible only in regions that elicit an evoked response to an experimental stimulus [7]. Both issues of properly detecting evoked activity and estimating the HRF then play a central role in fMRI data analysis. They are usually dealt with independently with no possi- ble feedback although both issues are strongly connected one to an- other. To introduce more flexibility regarding the assumptions on the HRF model, a novel approach referred to as the Joint Detection Es- timation (JDE) framework has been introduced in [1] and extended in [2] to account for spatial correlation between neighboring voxels in the brain volume (regular lattice in 3D). In this latter approach, the HRF can be estimated while simultaneously detecting activity, in a region-based analysis, that is on a set of pre-specified regions also named parcels partitioning the whole brain data set. This approach is mainly based on: (i) the non-parametric modelling of the HRF at a regional spatial scale (parcel-level) that provides a fair compromise
En savoir plus

5 En savoir plus

Bayesian joint detection-estimation in functional MRI with automatic parcellation and functional constraints

Bayesian joint detection-estimation in functional MRI with automatic parcellation and functional constraints

3.4 Conclusion In this chapter, we introduced the JPDE model which is an extension of the parcel-based JDE model. This model assumes that a single unknown HRF shape is driving hemodynamic responses in a given parcel. Activated voxels within the parcel are then localized by inferring a spatially regularized bi- linear model. One major limitation of the JDE model that it requires the parcellation to be fixed a priori by, e.g., using clustering algorithms. The JPDE model solves this problem by avoiding the pre-defined parcellation. This model allows the grouping of the regions that share a similar HRF pat- tern and relaxing the hard constraint of a single HRF profile over a given parcel to cope with possible parcellation errors. These concerns were ad- dressed by introducing HRF patterns represented by Gaussian distributions and assigned to representative voxels using latent variables. These latent variables are governed by a hidden Markov random field, namely a Potts model that enforces spatial correlation between neighbouring voxels. How- ever, the number of the hemodynamic territories (parcels) has to be specified a priori for the JPDE model. This number has a huge influence of the detec- tion and estimation tasks and its adjustment is generally a non-trivial task. In this context, we proposed a variational model selection procedure based on the free energy calculation. This procedure was added as an extension to the JPDE model where the free energy was calculated for different candidate models after convergence. Each of these models is characterized by a given number of parcels and the model maximizing the free energy is the best fit for the fMRI data. In other words, if we have Ω different models then we have to run the JPDE model with the model selection procedure Ω times and compute the value of the free energy each time. The proposed extension was validated using synthetic and real data experiments. For synthetic data, the proposed procedure managed to estimate the correct number of parcels (when compared to the ground truth) for all the experiments. As regards real data, the region of interest was the temporal lobes and the model with two parcels was selected as the best data fit. These results are coherent with those obtained by the JPDE model in ( Chaari et al. , 2012 ) where two similar HRF profiles were estimated in the left component and one HRF profile was estimated in the right component.
En savoir plus

232 En savoir plus

Show all 10000 documents...