Neuroimaging

The following literature review involves both Vision and Machine Learning (ML) areas, focusing not only on understanding how subcortical regions of the human brain interact to perform specific cognitive tasks, but mainly on how such understanding can be used for task prediction using vision tools and pattern analysis. Vision techniques are usually applied for low level preprocessing stages, including spacial and temporal filtering and denoising but some higher level procedures are used as well, such as registration, interpolation and segmentation. ML techniques include Correlation Based Classifiers, Linear Discriminant Analysis, Principal Component Analysis, Neural Networks, EM based (iterative) Linear Regression and Linear Support Vector Machines among other (mainly linear) classification algorithms.

Reading

Ting, J.; D'Souza, A.; Yamamoto, K.; Yoshioka, T.; Hoffman, D.; Kakei, S.; Sergio, L.; Kalaska, J.; Kawato, M.; Strick, P.; Schaal, S. "Predicting EMG Data from M1 Neurons with Variational Bayesian Least Squares", in: Weiss, Y.; Schölkopf, B.; Platt, J. (eds.), Advances in Neural Information Processing Systems 18 (NIPS 2005), Cambridge, MA: MIT Press. [pdf]

Analyzing high dimensional data is a common task for many neuroscientific experiments, but linear regression like methods result impractical. This paper presents a EM algorithm that iteratively finds the linear regression coefficients and at the same time, builds a map that tells between relevant and irrelevant input dimensions for generalization. Performance evaluation is done in the context of predicting EMG (extracellular field potentials produced by muscle activation) from monkeys' neural data, or in other words, how muscles respond to a particular neural activity pattern. A performance baseline was obtained by using a brute force combinatorial analysis choosing a subset of neurons to predict the EMG data, the proposed algorithm performed as good. (Nov, 06)

Thirion, B.; Roche, A.; Ciuciu, P.; and Poline, J. "Improving Sensitivity and Reliability of fMRI Group Studies through High Level Combination of Individual Subjects Results", CVPRW 2006, IEEE Computer Society, Washington, DC, 62.

The idea of doing inference at higher levels rather than by simply analyzing voxels is exploited in this paper. The method first finds regions of interest (ROIs) defined as the set of connected components of thresholded fMRI images, then looks for close regions activated within different subjects and finally constructs a probabilistic correspondence map based on relative positions of these ROIs in order to build cliques of inter-subject corresponding regions. The (loopy) belief prop algorithm (used to build pair wise correspondences probability maps) is initialized by using distances between active regions of two individuals and messages are designed as to encourage similar spatial distributions of active regions (again, between both individuals). Results showed better performance when compared with state of the art technology, but methodology (connected components of thresholded datasets) does not seem to match what domain experts usually do. (Nov, 06)

Kalanit Grill-Spector, Rory Sayres and David Ress, "High Resolution imaging reveals highly selective nonface clusters in the fusiform face area", Nature Neuroscience, Vol. 9, No. 9, September 2006

Standard resolution fMRI (SR-fMRI) used in previous research showed that the Fusiform Face Area FFA is a set of clustered face-selective neurons given its stronger response to faces an weaker response to non-face objects. However, authors in this paper show that, by analyzing High resolution fMRI data, the FFA is composed by a discontinuous mixture of face-selective neurons, and neurons with strong responses to non-face objects as well. Authors also show that SR-fMRI average the response of this heterogeneous set of neurons but given the higher concentration of face-selective neurons, the averaged response could be associated to face selectivity.

Methodology involves a method that uses a few tenths of voxels to effectively tell between four different stimulus (faces, animals, cars and sculptures) given the observed activations. (Nov, 06)

Peter Savadjiev, Jennifer S. W. Campbell, G. Bruce Pike and Kaleem Siddiqui, "3D Curve Inference for Diffusion MRI Regularization", Neuroimage, October 2006

A geometrical approach that combines information from a fiber neighborhood to infer local fiber orientations using diffusion MRI data. The idea is to treat fibers as 3D space curves and, for each voxel, the calculated confidence on a particular orientation depends not only in its local (voxel) information, but in its compatibility with neighboring orientations. The effect is that co-circular fibers lend support between one another. A function called the "Average Local Support" integrates the confidence of each voxel and the neighboring compatibility, and this function is then maximized using a Relaxation Labeling technique. (Nov, 06)

Norman, K.A., Polyn, S.M., Detre, G.J. and Haxby, J.V. "Beyond mind-reading: multi-voxel pattern analysis of fMRI data. Trends in Cognitive Sciences", 10(9), 2006.

What is currently known about knowledge representation is still extremely vague, but increasing evidence shows that multiple brain areas interact during most cognitive tasks. Consequently, a multi-voxel pattern analysis (MVPA) is needed to describe (if possible) mental representations in the brain, rather than single voxel approaches, as this technique provides "increased sensitivity". Here, a number of papers where MVPA is naturally applied are listed, but also some issues about why to use a particular (linear, non-linear) classifier along with the difficulties to interpret weights of successful classifiers. This issue of "weight interpretation of a trained classifier" is not new to Machine Learning community but it is certainly not as studied as the learning itself, yet is key in most cognitive representation research problems. The last section summarizes a set of questions for future research, including of course "how to search the brain for voxel sets that help classification tasks?”. (Oct, 06)

Haynes J, Rees G, "Decoding mental states from brain activity in humans", Nature Reviews, Neuroscience, Vol. 7, No. 7. (July 2006), pp. 523-534.

The idea of "Multivariate Analysis" is confronted against the "Strictly location-based conventional analysis" motivated by the fact that weak information can be accumulated in an efficient way across many spatial locations and because several brain regions could potentially carry information about a single cognitive state (and maybe not when analyzed separately). This paper includes a sensible and educated description of the current technical and methodological challenges of developing effective neuroimaging technology, including issues such as spatial and temporal resolution, invariance, superposition extrapolation and generalization across individuals. Ethical issues are mentioned as well. (Oct, 06)