Traditional poster

Spatial Independent Component Analysis (sICA) is increasingly being used for the analysis of fMRI datasets with unpredictable response dynamics, like in resting state experiments. However, group-level statistical assessments are difficult, and proper statistical characterization and validation under the null-hypothesis are so far lacking. In the current study, a novel method is proposed that is based on retrospective matching of individual component maps to aggregate group maps. Selection bias is analytically predicted and explicitly corrected for. It is shown that valid outcomes are obtained, in the sense that the achieved specificity does not violate the imposed confidence levels, only if bias-correction is applied. Sensitivity and discriminatory power remain acceptable, and only moderately smaller than those of a biased method. Finally, it is shown that the method is able to identify significant effects of interest in an actual dataset, proving its applicability as a group-level sICA fMRI data analysis method.

We introduce a new assumption- and parameter-free method for the analysis of fMRI resting state data based on „eigenvector centrality”. Eigenvector centrality attributes a value to each voxel in the brain such that a voxel receives a large value if it is strongly correlated with many other nodes that are themselves central within the network. Google's PageRank algorithm is a variant of eigenvector centrality. We tested eigenvector centrality mapping (ECM) on two resting state scans of 35 subjects, and found a network of hubs including precuneus, thalamus and sensorimotor areas of the marginal ramus of the cingulate and mid-cingulate cortex.

Network analysis of resting state fMRI data requires the specification of ROIs. This is a difficult process fraught with error. We propose a method for developing an ROI atlas by whole brain parcellation of resting state data in functinally homogenous, contiguous regions.

We have applied a Bayesian non-parametric multiscale technique, the iterated Multigrid Priors method, to map the digits of the hand in primary somatosensory cortex for 1mm isotropic spatial resolution data. It is data driven and makes no assumption about the local hemodynamic response as a function of time or space. It was able to detect an orderly pattern of response phases on the posterior bank of the central sulcus (postcentral gyrus) suggesting that the method can also be extended for retinotopic mapping studies of visual cortex. We also showed variations in HRs across digits through local posterior spatial averages.

Support Vector Regression is a machine learning technique that learns the mapping from the training set and labels provided. This creates a model which can then be used to give predictions for all testing sets. The prediction is really quick and hence SVR has potential to be used as a tool for real-time biofeedback applications to evaluate graded potential. In this study, we have used SVR analysis to evaluate graded activation in multiple neural systems namely the visual and motor cortex activation. The outputs are encouraging and advocate prospects of using SVR for future work in building real-time biofeedback applications in which graded activation needs to be evaluated.

Support vector machines (SVM) and relevance vector machines (RVM) are two machine learning algorithms which have gained popularity due to its sensitivity to networks of brain activation. Despite their recent extensive use in fMRI research, little contribution has been put forth to compare these different algorithms. Both models were compared for speed and prediction accuracy. The results revealed that both RVM and SVM are comparable in classification accuracy. However, RVM is capable of performing the task much faster and with a sparser model. Feature selection was also found to increase both speed and classification accuracy for both SVM and RVM.

Propofol is an anesthestic agent widely used in clinical practice which is known to affect episodic memory. The exact mechanism causing this effect is still unclear. Here we investigated whether propofol has a region-specific effect on functional connectivity in fMRI data. Subjects were scanned under the influence of propofol or a placebo. Functional connectivity was assessed using an algorithm new to fMRI data analysis called 'eigenvector centrality'. Our results suggest that the well known impairment of episodic memory after propofol infusion is related to an impaired function of cerebellar regions known to be involved in memory encoding.

Data-driven analysis is useful in pharmacological MRI where there may be no model of neural response available a priori. It is recognised that the signal complexity of noise will usually be higher than any signal of interest. Renyi entropy may be used to discover the complexity of a time frequency representation of a voxel time course. Its application here at every voxel in a region of interest across several subjects shows it is capable of discovering drug effect which is not found when the same analysis is carried out on placebo data.

We used fMRI to assist single-trial EEG signal classification by transforming scalp EEG into corresponding source activation patterns. The classification performance for 4 categories visual perception task was around 98%.

Setting activation thresholds remains a challenge in functional MRI. While strategies exist to address the increased chance of false positive activations due to the large number of voxels in an fMRI image, these methods frequently ignore differences in activation strength between tasks, individuals, and scanners. Setting appropriate thresholds is particularly pertinent in presurgical mapping, as knowledge of the location and extent of functional cortex can affect surgical decisions. In this work, we demonstrate an automated threshold technique based on test-retest imaging and receiver-operator characteristic curves, which produces individualized threshold levels optimized for reproducibility of the observed activation.

In recent work we showed that by means of sparse estimation techniques the spatial and temporal evolution of single-trial BOLD responses can be automatically detected without any prior knowledge of the stimulus timing and without thresholding: paradigm free mapping (PFM). However, fMRI time series also contain physiological and instrumental fluctuations which can hinder the detection of BOLD responses associated to neuronal activity. Physiological fluctuations can be removed prior to PFM via high-pass filtering, or by RETROICOR, RVT or RVHRCOR, but these techniques must be employed in a pre-processing stage and require the additional recording of physiological respiratory and cardiac waveforms. Here, extending on our previous work, we present a novel technique which by decomposing the fMRI signal enables automatic detection of fMRI BOLD responses without prior stimulus information and automatic fitting of significant frequency fluctuations present in the signal, such as non-neuronal cardiac and respiratory fluctuations (semiparametric PFM, sPFM). This technique is based on a semiparametric linear representation of the fMRI signal which is recursively fitted using a morphological component analysis algorithm. The feasibility of this technique was evaluated in simulations and real fMRI data acquired at 7T, and its performance validated to RETROICOR.

A pattern-based rt-fMRI system capable of multi-session and group-based models enables progressive training and testing across sessions, and potentially enables the use of group models for rehabilitation/therapy using multi-voxel targets built from databases of recovered individuals. Here we investigate alignment strategies to verify that there is not a significant tradeoff between classification accuracy and rt-fMRI computational demands. Our results demonstrate the feasibility of a model-to-scan alignment system for real-time fMRI in which the least demanding computational approach does not lead to a compromise of classification accuracy. This work also demonstrates the feasibility of using group SVM models in real-time experiments.

Multivariate statistical analysis has recently become popular in fMRI data analysis as such methods can capture better the spatial dependencies between neighboring voxels. One such method is local canonical correlation analysis (CCA) where one looks at the joint time courses of a group of neighboring voxels. It is known that CCA without any constraints can lead to significant artifacts and an increase in false activations. Here, we investigate different novel linear constraints and a nonlinear constraint for CCA and propose a method that rectifies the weakness of conventional CCA mentioned above.

Functional sub-divisions of important anatomic regions in the human brain are normally done based on disparities in structural connectivity patterns or functional connectivity maps. However, quantification of functional heterogeneity, and determining the appropriate number of sub-regions, has rarely been a focus of study. This work evaluates the use of self organized maps (SOM) to classify the functionally different regions in the hippocampus, which exhibits functional and sometimes anatomical differences in patients with disorders such as schizophrenia and bipolar disorder etc. Using voxel based connectivity maps we successfully parcellated left hippocampus and found performance of SOM to be superior in comparison to kmeans clustering.

Presence of propagating spatiotemporal waves in low frequency fluctuations (LFFs) has recently reported using high temporal resolution single slice BOLD fMRI of the rat brain. We have developed a novel method for automatic detection of such patterns and some initial findings for multslice rat and human data are presented in this abstract.

The T2* contrast mechanism associated with the BOLD signal is well known, as is its echo time (TE) dependence. In this abstract, we present a method for analyzing data acquired with interleaved echo times. Based upon the expected BOLD TE behavior, the ratio of the regression coefficients for the task related columns of the design matrix may be used to identify voxels which exhibit BOLD-like responses.

We propose the eigenspace L1-norm beamformer, a new novel technique for ultrafast MR inverse imaging (InI) reconstruction. This method minimizes the amplitude of the beamformer output quantified by the L1-norm of the spatial filter coefficients. We tested this method to reconstruct functional MR InI measurements using a visuomotor task. Results show that the eigenspace L1-norm beamformer can detect BOLD contrast functional activity and provide higher spatial resolution than linear constrained minimum variance (LCMV) beamformer in both motor and visual cortices.

BOLD fMRI is restricted by low signal to noise and various artifacts varying from motion to physiological noise. Independent components analysis (ICA) is a data-driven analysis approach that is being used to filter fMRI of such noise. However, one of the problems with ICA remains the interpretation of the results. Recently, we developed an automatic classifier (Spatially Organised Component Klassifikator - SOCK), which uses spatial criteria to help distinguish plausible biological phenomena from noise. We utilize SOCK to automatically filter a conventional fMRI block-design language study and successfully show the significance of activation obtained increases as a result of SOCK.

The first applications of real time fMRI used voxel intensity averaging over a ROI to provide feedback, whereas recent work has shown that machine learning tools may improve performance. We conduct a comparison between the two techniques, and find that support vector machines (SVM) outperform averaging over a ROI no matter how restricted an ROI we use. Further to this, we find that SVM performance does not decrease as sharply as ROI averaging when block length is decreased.

Dynamically Adaptive Imaging (DAI) is a new real-time paradigm for fMRI. BOLD data were analyzed using our open-source real-time software and used to iteratively and automatically adapt the stimuli presented to the volunteer. DAI was applied to investigate feature coding in ventral visual cortex. Pictures of objects were presented on a screen. We performed an iterative search, in which the outcome of the experiment was the neural neighborhood of stimuli that evoked the most similar pattern of neural response to a referent stimulus. DAI converged rapidly and found object-specific tuning to complex conjunctions of sensory and semantic features.

Combined EMG and fMRI is very desirable. Detecting changes in muscle activity associated with changes in cortical activity can greatly improve our understanding of neuroplastic changes and the affects of treatments in neuromuscular conditions. This work proposes a robust wavelet-based artifact reduction strategy that allows for the distinction between two muscular conditions in an MR environment. This work also introduces the use of the EMG parameter median frequency as a covariate in a motor fatigue study to better refine image analysis.

Resting-state fMRI analysis techniques that determine the similarity between time varying signals of seed and target regions assume the signals are stationary; however, the resting-state varies between subjects and is susceptible to unwanted brain activity due to inadvertent movements or cognition. In this study, we introduce a time-frequency approach based on the Stockwell transform to temporally resolve coherence between resting-state signals. We demonstrate S-Coherence can reduce the contribution of unwanted hand movements in the determination of the resting-state connectivity within the motor network, and hence reduce within-subject variability in comparison with existing techniques (temporal cross-correlation and coherence).

Here we present an olfactory fMRI data processing method that can significantly improve the data processing quality when the patients’ respiration pattern is not controlled and doesn’t synchronize with olfactory stimulation paradigm. As an example of implementation we present an olfactory fMRI examination while the subject’s respiration pattern is not regular. Our data demonstrates that it is critical to consider the subject’s respiratory patterns’ modulation on the olfactory stimulation paradigm. The presented olfactory fMRI data processing method can be used for various applications. In addition to the example of real time respiration data, subjective response data (not provided here) can also be convolved with odor delivery data for more improved fMRI data processing. This experimental set-up will be useful in the olfactory fMRI study of neuropsychiatric and neurologic patients that are not cooperative or be able to follow the breathing instructions.

Physiological noise caused by ecg- and/or breathing related pulsatility may introduce temporal correlations that are unrelated to neuronal processes in a resting-state network analysis. As physiological noise is often non-linear and non-stationary, signal extracted by simple filtering will deviate from the actual noise, and so as global regression methods like RETROICOR. In this study we implemented empirical mode decomposition (EMD) on resting-state fMRI time-series and extracted cardiac components which has a time-frequency curve that well matches the true heart rate acquired by external ECG during the scan.

3D FMRI acquisitions have the advantage of allowing high resolution, isotropic, imaging. However, 3D acquisitions, such as SSFP and SPGR, show increased signal instabilities in the inferior regions of the brain. Here, we present a characterization of these temporal instabilities and propose a GRAPPA-based correction method that allows retrospective gating of 3D FMRI data.

fMRI analyses are primarily based on the magnitude information in gradient-echo echo-planar images (GE-EPI) but a growing number of studies also included the phase information. An issue relates to physiologic large-scale phase effects that are more prominent in phase than magnitude data. In the present work we explored the phase stability at different length scales at 3T and found that improvements in temporal stability could be achieved by alternative noise-reduction methods that take into account the differential origin of noise effects in phase and magnitude data.

High dimensional feature space of fMRI volumes has been a drawback for classification studies since large feature dimension is known to increase the classification error and the computation time. In this study, we combined PCA with two anatomical feature selection methods: grey matter (GM) and region of interest (ROI) masking, and investigated the effects of different feature reduction methods on the classification accuracy of a linear SVM classifier. To apply PCA after anatomical masking is concluded to be a reliable method for preserving the classification accuracy of the anatomical feature selection methods and reducing the computation time.

EPI based fMRI has several major limitations: distortion, low spatial-resolution, and low anatomic resolvability. Therefore, it is not easy to register fMRI data to structural images, and to normalize fMRI data. EPI distortion correction and nonlinear normalization methods have been developed to address these limitations. However, it is not easy to assess how these methods perform on fMRI data with distortions, limited resolution, and anatomic resolvability. Here we report an imaging protocol based on high-resolution inversion-recovery prepared segmented EPI (with identical distortion patterns as in single-shot EPI), enabling accurate assessment of the performance for distortion correction and nonlinear normalization algorithm.