The Section on Functional Imaging Methods (SFIM) continues to develop cutting edge methods at the interface of functional MRI (fMRI) technology, image processing, interpretation of the signal, and neuroscience applications. The goal of SFIM is to fill the much-needed gap of improving fMRI from both a methodological and usability standpoint.
We aim to increase the depth and breath of fMRI applications and bridge the gap to clinical applications on individual patients. This report describes progress in our research. In the past year, we have published 16 papers, making this one of our most prolific years. Feature Selection and classification accuracy. Carried out by my post doc Carlton Chu. A growing area in fMRI is the use of Machine Learning algorithms to characterize fMRI patterns of activity as they are associated with specific tasks. A central problem in this approach is regarding how to chose ahead of time, the appropriate region on which to perform these calculations. Many classification approaches simply use all voxels in the entire brain. While this works, it includes many voxels that are not involved with the task itself. This paper addresses this issue and determines that, in fact, feature selection is better, yet depends a bit on the algorithm being used. Multi-echo acquisition and Independent Component Sorting Carried out by my graduate student Prantik Kundu as well as E. Bullmore, N. Brenowitz, J. W. Evans, S. J. Inati, W.-M. Luh, V. Roopchansingh, Z. S. Saad. An ongoing challenge has been to remove the large fraction of non-neuronal fluctuations rom rs-fMRI time series. Work on characterizing and removing non-neuronal fluctuations has focused on time series modeling based on external measures of physiologic processes. In this study, we use the TE-dependence to separate BOLD from non-BOLD time signal. BOLD signal changes are manifest as changes in T2*, which can be characterized as showing a linear increase in fractional signal change with echo time (TE). Motion, system instabilities, and inflow effects can be manifest as changes in longitudinal relaxation, T1 or proton density, S0 but not typically T2*. Muti-echo for removal of systematic motion effects in group comparisons Carried out by my graduate student Prantik Kundu. It has been shown that in-scanner subject head motion leads to systematic patterns of false positives in population-level rs-fMRI analysis. This artifact is worst for group contrasts of cohorts with different levels of in-scanner motion, and is a potential roadblock to further resting state fMRI - based research. Here we show that me-ICA enables elimination the problem of spurious functional connectivity (FC) patterns caused by motion and other artifacts. After decomposing multi-echo data with ICA and separating BOLD from non-BOLD components based on echo-time (TE) dependence, the BOLD-based ICA components are used to compute FC. Computing an individual-subject me-ICA functional connectivity map for a region of interest (ROI) involves computing the correlation of its ICA series with the ICA component series of all other voxels. In contrast, the typical approach to computing FC maps involves motion regression, band pass filtering, and then Pearson correlation. The me-ICA seed-based FC approach was compared to standard seed-based FC approach at the population level. Thirty-three subjects were scanned for 10 minutes EPI during rest (Siemens 3T Tim Trio, 32-channel). No increase in me-ICA FC false positive rate is found as groupings are increasingly biased according to motion. We conclude that me-ICA FC is a principled and straightforward solution to the problem of spurious FC due to motion artifact. Clustering of resting state correlations. Carried out by Prantik Kundu. A major goal in rs-fMRI is the whole-brain mapping of functional cortical networks down to the scale of voxels from a single dataset. Promising results have been obtained on highly subject-averaged data, however, our ultimate goal is to use the characteristics of these networks as individual biomarkers. Towards this goal, we have developed an approach, based on our me-ICA method that converges on stable and consistent areas with very few averages. The maps that we produce can also predict subject-specific activation patterns associated with activation-based paradigms. This novel implementation of hierarchical clustering was able to create a community structure of brain organization, showing how large modules at low levels of clustering are related to smaller modules at high levels of clustering. We found that the 30-50 BOLD components from individual subject high dimensional me-ICA have localization to specific, finely delineated functional areas. The ability to consistently identify functional areas without functional localization tasks could also be of value as a biomarker for individual differences or disorder characterization. Determining most and least stable cortical networks using resting state fMRI Carried out by my post doc Javier Gonzalez-Castillo. An assumption has been that connectivity patterns are stable for the duration of the scan. The purpose of this work is to characterize the temporal variation of resting state connectivity within a continuous 1 hour resting state scan. From this we are able to determine the most and least stable networks. We compared pairs of matrices that were created by averaging the signal over varying amounts of time ranging from 2 to 30 minutes. The correlations begin to show decreasing similarity below 10 minutes, and continue to drop sharply down to 2 minutes. We used a sliding window correlation analysis with a window duration of 1 min and a window step of 1 s. We then sorted the pairwise connections from most to least variable. While some connections remain quite stable across time, others seem to vary considerably. It can be observed that stable connections tend to by inter-hemispheric and symmetric. The unstable connectivity patterns correspond to these between subcortical regions and high order cognitive regions in frontal and predominantly left parietal cortex. Decoding """"""""yes"""""""" from """"""""no"""""""" answers. Carried out by Z. Yang. Towards the goal of determining the limits of sensitivity of the fMRI decoding approach, our goal was to determine of we could differentiate a simple yes or no response or subjective correctness, based on subjects fMRI activation patterns as they responded to simple common-knowledge questions. In each trial, we present a cue to instruct the subject to either honestly or dishonestly answer the following questions. After reading the questions, subjects have to keep their final answers in mind until a button-press prompt appears several seconds after the question are removed from the screen. Looking at the BA9 portion of the left middle frontal gyrus. We extracted five time points of the voxels within this region, starting at the onset of the questions. Treating each voxel as a feature and each time point as a sample, we trained Gaussian Nave Bayesian (GNB) classifiers to predict the truthful Yes/No answers to the questions. The mean prediction accuracy achieved 90% when we averaged all 20 trials. These findings indicate that the truthful answers are encoded in brain activity independently from intentions, and multivariate pattern analysis is able to decode them from fMRI signal.
Showing the most recent 10 out of 84 publications
