Workshop 6: Brain Imaging: Abstracts and Lecture Materials
Binocular vision is central to both space and form perception and plays a critical role in visuomotor feedback control. In this talk I will describe work that demonstrates a dissociation of perception from feedback control in a way that highlights the different computational requirements of these tasks. Attempts to look at the cortical loci of these processes will be described.
Quantitative analysis of brain structure and function is important in the study of many neurological and neuropsychiatric disorders. This talk will present work grounded in the use of spatial constraints and mathematical optimization to analyze neuroanatomical structure and function of the human brain from Magnetic Resonance Images (MRI). We will first describe our approach to segmenting cortical gray matter using a coupled level set strategy. Next, we will present an approach to subcortical segmentation based on the use of both object self-shape and neighborhood spatial relationship priors, both embedded in a level set- parameterized, maximum a posteriori (MAP) estimation framework. Finally, we will discuss recent work aimed at incorporating prior knowledge of brain activation patterns and segmented anatomical information (gray matter/white matter) to provide improved estimates of activation strength in a functional MRI (fMRI) attention-modulation experiment, again using a MAP estimation approach.
Behavior is assumed to emerge from specific circuits in the brain. These circuits are routinely inferred from functional brain imaging patterns. Differences in patterns of functional images between, for example, task conditions, drug conditions, and between control and pathological conditions are routinely used to inform researchers of basic biological mechanisms and pathophysiological processes in normal and abnormal brain function. There are, however, multiple levels and principles of organization of brain circuitry, often beyond the resolution and/or functional capabilities of imaging techniques such as PET, fMRI, and DTI. Furthermore, each neurological/psychiatric disorder differentially affects neuroanatomical modules and types of circuitry, and these must be borne in mind in the analyses and discussion of implied circuitry in imaging experiments.
We present iCluster, a fast and efficient algorithm that clusters a set of images while co-registering them using a parameterized, nonlinear transformation model. The output of the algorithm is a small number of template images that represent different modes in a population. This is in contrast with traditional, hypothesis-driven computational anatomy approaches that assume a single template to construct an atlas. We derive the algorithm based on a generative model of an image population as a mixture of deformable template images. The experimental results demonstrate that the algorithm can discover interesting sub-populations, suggesting applications in atlas-based segmentation and statistical analysis of anatomical differences in clinical studies.
This is joint work with Mert Sabuncu and Serdar Balci.
Functional MRI (fMRI) not only allows for localizing specific brain activity but also provides temporal information that can be used to derive connectivity within the brain. This paper will describe some techniques for assessing brain connectivity and demonstrate them with exemplar applications. Specifically, we will discuss assessing local coherence in fMRI data based on the integration of correlation function and its neurobiological relevance and applications. In addition, analysis of effective connectivity using Granger causality analysis will be described, with an emphasis on a multivariate approach and its application to examine group differences in tactile perception and to examine dynamic changes in the motor network during fatigue.
Many frameworks have been proposed for the analysis of brain DW-MRI imagery. The objective of these frameworks is to yield a greater understanding of structure and connectivity within the brain and the relation of these to function. In this talk, we present a framework for the analysis of DW-MRI datasets that consists of two components: 1) an optimal path connecting two regions of interest and 2) a volumetric fiber bundle segmentation, initialized on the optimal path. This framework has the advantage of providing both connectivity and structural information about fiber bundles. Also, in this talk, we discuss the pros/cons of this framework and challenges in the state-of-the-art of fiber bundle segmentation.
Computational Anatomy is the study of the shape and structure of manifolds in human anatomy. This talk reviews results from CA along these lines, including (i) embedding of shapes into a metric structure via flows of diffeomorphisms (ii) conservation laws for geodesics describing metric connection of shapes (iii) statistics on families of shapes encoded via these metrics. The emerging focus in Computational Functional Anatomy is the inclusion of the study of function in the curved coordinates of anatomical manifolds. Methods for performing inference in this setting are examined coupled to morphometric studies.
Automatic segmentation of pathological brain MR images with tumors is crucial for analyzing tumor characteristics, efficacy of drug treatments, and surgical planning. Many segmentation schemes have been developed, yet validation and performance comparisons are difficult since there are no public brain tumor MRI databases with consistent, reliable ground truth. In this talk, we will present the practical use of statistical and physical modeling for generating synthetic brain tumor MR images with known, objective ground truth. We combine a physical deformation model and a physical infiltration model to generate anatomical data with pathological structures (tumor and edema). We then use a statistical image generation model to obtain synthetic multi-modal brain tumor MR images that correspond to the generated anatomical data. The synthetic brain tumor MRI database has potential uses for validating different segmentation schemes, for surgical simulations, and for clinical training.
Orientation distribution functions (ODF) can be used to represent multiple fiber crossings in the brain as recorded by High Angular Resolution Diffusion Imaging (HARDI). Current state-of-the-art methods use spherical radial basis functions or spherical harmonics to represent the ODF. These methods however require many coefficients to represent each ODF and ambiguities can occur when the principal diffusion directions are to be extracted. In this talk, we propose to use "directional functions" for representing the signal and provide closed form expressions to approximate the corresponding ODF. These functions require very few parameters (three) to represent the ODF and the principal diffusion directions are naturally obtained during the estimation process. We will also show how to perform interpolation using these directional functions and propose 2 metrics, a Euclidean and a hybrid-Euclidean-Riemannian, to compute geodesic distances between 2 ODFs.
In this talk, we provide a framework for extracting tubular structures from medical imagery. The general methodology will be applied to modeling and extracting the cingulum bundle (CB) from diffusion-weighted imagery (DW-MRI) of the brain. The CB is a tube-like structure in the brain that is of major importance to clinicians since it may be helpful in diagnosing schizophrenia. This structure consists of a collection of fibers in the brain that have locally similar diffusion patterns, but vary globally. Standard region-based segmentation techniques adapted to DW-MRI are not suitable for this application because the diffusion pattern of the CB cannot be described by a few simple global statistics. Typical active surface models extended to DW-MRI allow for arbitrary deformations that give rise to unlikely shapes, which do not respect the tubular geometry of the CB. In this work, we explicitly model the CB as a tube-like surface and construct a general class of energies defined on tube-like surfaces. Modeling the CB as a tube-like surface is a natural shape prior. Since a tube is characterized by a center-line and a radius function, the method is reduced to a curve evolution that is computationally much less costly than an arbitrary surface evolution. Our tubular model of the CB also has the advantage that computing shape statistics and functions defined on the CB are simplified.
Segmenting cortical and sub-cortical structures from 3D brain images is of significant practical importance. In this talk, we will discuss a new statistical modeling/computing framework and show its application for whole brain segmentation. The notion of using context information for solving the medical imaging problem has been increasingly realized in the field. However, how to learn an effective and efficient context model, together with the image appearance, remains mostly unknown. The current literature using Markov Random Fields (MRFs) and Conditional Random Fields (CRFs) often involves specific algorithm design, in which the modeling and computing stages are studied in isolation. Medical images observe complex patterns, contributed by many factors such as textures (homogeneous, inhomogeneous, and structured) and machine parameters. This auto-context model is about a new attempt to push the appearance and context information in a seamless way by automatically incorporating a large number of short-range and long-range features. The resulting algorithm has nearly the identical procedures in computing (testing) as in modeling (training), and thus, achieves rapid performance the holistic medical image segmentation task. We will show a variety of sub-cortical and cortical segmentation results using this model.
Imaging as a biomarker of drug response is becoming an increasingly important field of research. Government, industry and academia have agreed to collaborate on improving the development of therapies and outcomes for common diseases, especially cancer, through biomarker development and evaluation. Biomarkers are biological indicators of disease or therapeutic effects that can be measured by in vivo biomedical imaging and molecular imaging in particular, as well as other in vitro or laboratory methods. Recent work has shown that biomedical imaging can provide an early indication of drug response by use of CT, MRI and PET/SPECT.
Many sources of uncertainty exist in imaging as a biomarker. Biological variability, for example, is a factor both drug- and patient-dependent and thus difficult to characterize or model. However, other uncertainties are associated with the image data collection platform and the robustness of software tools required for reliable, quantitative measurement of change over time, such as tumor volume, radioactive tracer activity, or contrast agent dynamics. All these sources of uncertainty significantly affect the statistical power of clinical drug or therapy trials.
The challenges and opportunities for imaging biomarkers are explored for brain imaging, especially brain traumatic injury and developmental disorders.
We formalize the pair-wise registration problem in a maximum a posteriori (MAP) framework that employs a multinomial model of joint intensities with parameters for which we only have a prior distribution. To obtain an MAP estimate of the aligning transformation alone, we treat the multinomial parameters as nuisance parameters, and marginalize them out. If the prior on those is uninformative, the marginalization leads to registration by minimization of joint entropy. With an informative prior, the marginalization leads to minimization of the entropy of the data pooled with pseudo observations from the prior. In addition, we show that the marginalized objective function can be optimized by the Expectation-Maximization (EM) algorithm, which yields a simple and effective iteration for solving entropy-based registration problems. Experimentally, we demonstrate the effectiveness of the resulting EM iteration for rapidly solving a challenging intra-operative registration problem.
This is joint work with Lilla Zollei and Mark Jenkinson.
This presentation emphasizes the mechanics, rather than the methods, from data to publication. I will present Matlab software (SurfStat) for the statistical analysis of univariate and ultivariate surface data using linear mixed effects models (fitted by ReML) and random field theory (RFT). SurfStat is intended for cortical thickness data on triangular meshes, but it will handle any triangulated surface data; the only requirement is that the triangulation scheme must be the same for all surfaces, i.e. the data must be registered to a common surface. Inference uses RFT for T, F, Hotelling's T2 and Roy's maximum root statistics. An attractive feature is the use of a model formula rather than a design matrix for specifying the linear model. It is fast, because everything is loaded into memory, permitting a truly interactive analysis, with no need for batch. Finally, off-the-shelf Matlab graphics are ready to publish.
