This work focuses on model based spatial estimators for detecting activation signals in fMRI. As shown before there are improvements to be gained in exploiting the local correlation structure in the MRI signal. Both in terms of statistical power and accuracy. Besides PCA, a newer method which also attempts to find significant projections of the data, independent components analysis will be used. Kernel based methods, such as radial basis functions or flexible shape bases will be used to improve detection probability in local regions; given stronger assumptions about local coherence the detection accuracy should increase although with some tolerable increase in false positives. Such location estimators are stronger then the local correlation approaches because they make stronger assumptions about the underlying topology (shape).
Are the residuals Gaussian? Hanson & Bly have argued that the skewed, heavy-tailed nature of the distribution of blood oxygenation level dependent (BOLD) data does not suggest a Gaussian random noise process. We investigated this question further by examining residuals with different models for the autocorrelation structure to account for systematic noise, and this suggests that the residuals are non-Gaussian. It has been argued that the the random noise process for BOLD data may be Gaussian but heterogeneous. Preliminary examination of BOLD data suggests that the individual distributions at each voxel are overwhelmingly skewed, much more so than Gaussian samples. Hence we propose an alternative explanation -- that BOLD data have a common but non-Gaussian distribution.
Related Papers
This work seeks to capitalize on the spatial
resolution of fMRI image data and the temporal resolution of ERP (event related
potentials) data by exploring ways in which the two types of data can be
productively combined. Several new methods that have promise have been
developed and have been reported on at Signal Processing conferences (NFSI).
Shown in the figure is the constraints on the
fusion problem. Two forward models indicate the EEG and fMRI temporal and
spatial properties. Fusion on the left of this figure is defined by a derived
measure that is optimized with Gradient estimates or Linear Programming
methods.
Recent changes to the RUMBA software improvements in rumba-python Several improvements have been made to the python layer of the rumba libraries. These include the rumbagui execution framework, which is now implemented in rumba-python, and the migration of some previously unreleased GLM code into the rumba-python layer. rumbagui The rumbagui package has been redesigned with a new architecture that is based on a shared-memory model. The first release of the new version of rumbagui was demonstrated at the 2003 Human Brain Mapping conference. Motion Correction and Realignment Motion correction and realignment have been added to the rumba software. The related functionality includes between modality and within modality registration. Rigid body and 12 parameter models are used for within-subject fits, and polynomial transforms of degree up to 5 are used for between-subject or subjecttemplate fits. Coordinates in the template can be then mapped to Talairach space and used to extract Talairach labels. Further work has begun on regularisation for the polynomial based fits, to prevent the algorithm from over-fitting data. Lognormal parameter estimation A sample implementation of the algorithm used to compute the lognormal model in the article below has been added to estimate parameters.
Exploratory Data Analysis and the RUMBA Software Package
Data analysis in neuroimaging is a complex task which requires a large degree of automation via computer software. There are a number of tools that implement particular analysis strategies, such as SPM, FSL, AFNI, and VoxBo. At the other end of the scale, there are also general purpose scientific computing tools such as Matlab and other matrix computation packages that implement high performance general purpose matrix algebra, which is usually needed to implement the algorithms and statistical methods pertinent to the analysis of MRI data. We have implemented a middle ground to this software dichotomy in creating a library of tools to facilitate exploratory data analysis which is necessary in order to accomplish exploratory analysis such as the methodological research on the residuals described above.
An Overview of the RUMBA Tools
The RUMBA software has several components:
librumba
is a C++ library of reusable functions that implement functionality that is critical to the analysis of neuroimaging data. These include functions for tasks such as:
rumba-python
is a python module that makes the C++ functions accessible through the python
programming language. This provides a more high-level, interactive interface to the RUMBA
software. Python is not only interpreted, it can be used as an interactive shell and
rapid application development environment.
command-line programs
that are based on these C++ and Python functions
rumba-gui
that acts as a graphical script-builder,
offering a clean and intuitive way to present the
otherwise complex topic of function composition and operations on functions.
Recent Changes to the RUMBA Software
rumba-python
Several improvements have been made to the python layer of the rumba libraries. These include the rumbagui execution framework, which is now implemented in rumba-python, and the migration of some previously unreleased GLM code into the rumba-python layer.
rumbagui
The rumbagui package has been redesigned with a new architecture that is based on a shared-memory model. The first release of the new version of rumbagui was demonstrated at the 2003 Human Brain Mapping conference.
motion correction and realignment
have been added to the rumba software. The related functionality includes between modality and within modality registration. Rigid body and 12 parameter models are used for within-subject fits, and polynomial transforms of degree up to 5 are used for between-subject or subject-template fits. Coordinates in the template can be then mapped to Talairach space and used to extract Talairach labels. Further work has begun on regularization for the polynomial based fits, to prevent the algorithm from over-fitting data.
lognormal parameter estimation
A sample implementation of the algorithm used to compute the lognormal model in the article below has been added to estimate parameters.
Future Plans for the RUMBA Tools
realignment and motion correction We plan to implement new motion correction techniques based on stochastic gradient method. We also plan to improve the performance of our existing registration methods by incorporating regularization to prevent over-fitting.
rumba-gui
Developing, refining, and releasing a public beta of rumbagui is another short term goal. Longer term goals include support for parallel computing by allowing jobs to be processed asynchronously, a distributed client/server version that operates via a remote java application.
RUMBA tools History
6/5/01 RUMBA WEB site and distribution of RUMBA software available (www.rumba.rutgers.edu)
7/12/01 Coordinating with fMRID at Dartmouth. Investigating usage of RUMBA tools with fMRID
Related Papers
This set of experiments examines the neural correlates of bottom-up and top-down processing in the parsing of action sequences. Subjects watch video clips of action sequences and decide where event changes have occurred. This requires integrative neural circuitry across many areas of the brain.
Perceptual Cycle
Niesser originally described this type of integrative function as the perceptual cycle, in that it seemed to recruit encoding, filtering, spatial modeling and working memory function in some sort of feedback loop. The "perceptual cycle" can be seen to be comprised of three kinds of computational problems:
We will examine these questions as well as others by looking at the brain basis for schema activation, schema encoding, and schema boundary detection. Preliminary work implicates, areas such as cingulate gyrus, parietal areas, insula, previously related to attentional networks and decision related function such as dorsal lateral prefrontal cortex. We are interested in the network of relations between these areas and the way in which they are sequentially activated.
Structural Equation Modeling
We have recently started structural equation modeling of the covariance-variance matrices of neuroimaging data. This is similar in spirit to the work of Randy McIntosh, except we will start with a set of ROIs (regions of interest) and then search all possible graphs associated with all possible connectivity in the N! Space (for 9 ROIs is this 500k graphs). So far analysis has shown that the best fit graphs are SMALL (10s to 100s), meaning that other constraints could in fact help identify specific graph structure in a large set of regions.
Displayed below are examples of two kinds of stimuli used for the event perception studies.
On the left we have a "TOP-DOWN" example of a video which shows a actor studying. On the right-hand side we have a "bottom-up" example of a video, which shows a set of moving geometric shapes.
Preliminary work shows to sets of overlapping ROIs in different networks that seem responsive to the two kinds of stimuli. On average, the TOP DOWN type video seems to produce MORE areas, however the specific networks of areas will be our focus in these studies.
Related Paper
Zaimi, A., Hanson, C., & Hanson, S.J. (2004). Cognitive Neuroscience Annual Meeting, San Francisco, CA.
Subjects were asked to make similarity judgements about pairs of schematic faces based on the similarity of features (eyes or mouth) or the emotion expressed by the faces. Following the similarity task, subjects were asked to categorize new faces into two categories. For one group of subjects, the faces were categorized on the basis of a single feature (linearly separable task). For the other group, the faces were categorized on the basis of two features that varied orthogonally (integral task).
Experiment 2 replicated this design in an Allegra 3T magnet. Orientation toward emotional expression during the similarity judgement task significantly improved performance on the integral task but had no effect on the linearly separable task. We are interested in the nature of learning and attention in human behavior. Humans learn to categorize stimuli based on features of those stimuli. However, often category decisions depend on combinations of features (integral) rather than single (separable) features. The question we are interested in is what is the basis for integral and separable features as they affect learning and the nature of the brain mechanisms that might be involved in their perception.
We expect EEG and fMRI's studies during the above experiment, will produce different modulation of elements of several possible attentional networks (whose elements activates the following brain areas: caudate, insula, and possibly precunneus). What we would look for using both detection analysis (SPM) and network style methods (McIntosh) would be for both the amount of modulation of each putative attentional area and the strength of the links between the various areas--which may be diagnostic of separable or integral processing. These results could also be confirmed in a neuropsychological context with patients with neurological deficits in those regions and see the effects that arise when examining integral vs. separable stimuli.
Related Paper
Haxby et al. (2001) recently made an observation that the responses in ventral medial temporal
lobe to object identification were overlapping and distributed in topography. This observation
was in contrast to the prevailing view that object codes were focal and localized to specific
areas such as the fusiform and parahippocampal gyri.
We conclusively test this hypothesis and rule out the other two logical possibilities
(localist codes or unique distributed codes) that were consistent with the Haxby (2001)
analysis. Using a neural network classifier which achieves 85% generalization error across
all categories/subjects and a voxel-wise sensitivity analysis we show that the response in VMT
is combinatorial, that is, substantially the same voxels are contributing to the classification
of all visually presented objects. The results of eight-way neural network classifier are shown
in the figure.
Moreover, there appears to be no local representations useful for category assignment.
The neural network representations (hidden units) of the voxel codes are also shown to be
sensitive to each category, and for the first time in these types of neuroimaging data,
to a superordinate level feature (animate/inanimate) which was only available implicitly
in the object categories.
Related Paper
This research looks at how similar processing at encoding and retrieval affects implicit and explicit processing of information.
There have been several studies comparing computational models of categorization. Although, these comparative studies provided information on the modelsM-bM- capabilities or incapabilities of reproducing human-like categorization behaviors, they did not necessarily provide information that can lead to understanding of the nature of human category learning because of the confirmative characteristics of the models. In other words, model-to-model comparisons are not informative for testing plausibility of each assumption or theory of human category learning independently, rather it only allows omnibus tests collectively comparing all assumptions at a time. Since, it has been difficult to understand which assumption(s) is more likely to be tenable from the results of the previous comparative studies, it seems necessary to have a framework for modeling human category learning that allows us to manipulate and test a limited number of model assumptions at a time. The framework should be general and flexible, so that we can conduct standardized exploratory modeling of various types of human cognitive processes associated with categorization. The GECLE (for Generalized Exploratory models of Category LEarning) is a general and flexible exploratory modeling approach for human category learning that is capable of modeling human category learning with several different model assumptions. Furthermore, it allows manipulation a limited number of model assumptions at a time. For example, one can manipulate assumptions on how stimuli are internally represented (e.g. exemplars vs. prototypes) and or how people selectively pay attention to input feature dimensions (e.g., paying attention to dimensions independently or dependently). (Please check http://www-psych.rutgers.edu/~matsuka/tm_rumba.htm for detail)
Figure 1: Different types of GECLEM-attention mechanisms. Top row: shape and directions of receptive fields of all reference points are the same. Bottom row: a receptive field of each reference points is unique. Left column: all receptive fields are circular. Middle column. receptive fields are scaled orthogonally. Right column: receptive fields are scaled non-orthogonally.
Many neural network models of human category learning use a form of gradient algorithm. These methods have been successful in reproducing group learning curves, but tend to underpredict variability in individual-level data (Matsuka, 2002). In addition, many recent models of categorization have been criticized for not being able to replicate rapid changes in categorization accuracies and attention processes observed in the empirical studies (Macho 1997; Rehder & Hoffman, 2003). In the present study we introduced a stochastic learning algorithm for NN models of human category learning and show that use of the algorithm can result in (a) rapid changes in attention processes, and (b) different learning trajectories and more realistic variability in individual learning curves, especially for attention allocation.
Example results:
Figure 2. Predicted attention processes by stochastic learning algorithm showing capability of rapid shift in attention processes (bottom panel).
Figure 3. Top row-Empirical data, from left to right: classification accuracy, dimensional attention allocation, proportions of attention allocated to Dimensions 1 vs. 2. Middle row: Predictions by ALCOVE with gradient-based learning algorithm. Bottom row: Predictions by ALCOVE with stochastic learning algorithm