Baba Vemuri - US grants

Computer Vision Research Instrumentation will be provided for researchers at the University of Florida for research in the Department of Computer and Information Science. This equipment is provided under the Instrumentation Grants for Research in Computer and Information Science and Engineering program. The research for which the equipment is to be used will be in the areas of scene information, algorithm development for spatial reasoning, and algorithm development for fast access and retrieval of multidimensional, multisensory data.

This project will aim to establish a quantitative framework for integrating multi-sensory visual data in computer vision. Using the methods of variational calculus, the information from multiple visual sources will be embedded to achieve the goal of object surface reconstruction and recognition. In robotic applications spatial reasoning is important. Spatial reasoning means vision, task planning, navigation planning for mobile robots, symbolic reasoning and the integrating of such reasoning with geometric constraints. To the extent spatial inferences must be made from information gathered from multiple sources the ideas developed in this project on integrating information sources are an important adjunct to spatial reasoning. Since the formulation of the problem favors concurrent computing, both computer vision and spatial reasoning will benefit from this research.

9210648 Vemuri The goal of the proposed research is to develop a unified computational framework for several low-level vision problems which fall under the generic descriptions. Formulations in literature of a majority of these problems lead to minimization of non-convex functionals. Existing minimization techniques (stochastic or deterministic) are either computationally tardy or are efficient only under certain restrictive assumptions. Hence, there is a critical need to examine alternate optimization techniques that are not susceptible to pitfalls of the existing techniques, and the proposed research is an attempt in this direction. This research is concerned with the application of a relatively new technique called genetic algorithms (GAs) to a variety of visual reconstruction problems namely, stereo matching, discontinuity preserving surface reconstruction, and structure form motions. The proposed research will focus on issues involved in the analytical modeling of the GA using Markov chains to facilitate convergence analysis of the algorithm when applied to Visual Reconstruction problems. The theoretical work will be concluded with algorithm implementation and testing on real image data. The proposed unified computational framework will significantly advance the state of the art in computational vision. ***

DESCRIPTION (Adapted from Applicant's Abstract): Hippocampal asymmetry measure using Magnetic Resonance Imaging is a sensitive index of the lateralization of temporal lobe epilepsy. When the focus is confined to one hemisphere, surgery can provide considerable relief. Thus, measurement of hippocampal volume can provide information crucial for surgical planning. However, segmentation of the hippocampus remains non-trivial and is presently performed by laborious inefficient and difficult to reproduce manual procedures requiring expertly trained users. To date, no successful method is available to automate this procedure due to the poorly defined hippocampal boundary on MR images. Thus, a new approach is required in order to increase efficiency and predictive value of MRI for surgical planning gin the treatment of epilepsy. In this proposal, a potentially novel shape recovery method using a multi-resolution wavelet basis framework incorporating prior information on the hippocampus will be developed . The method will be trained using 90 retrospective clinical exams performed by an expertly trained neuroscientist. Preliminary data already demonstrate the great potential of this methodology. A user friendly Visualization and Menstruation Program (VAMP) will be developed as a clinical interface. When developed, VAMP will be tested retrospectively on 120 patient data sets not user for training the algorithm and the correlation between automatically determined hippocampal volumes and seizure frequency determined.

The goal of this research is the investigation of a scheme that allows for the representation of 2D and 3D shapes via compact and versatile geometric models which can model a large class of shapes and are amenable to stable and efficient numerical implementations. The models are to be capable of representing shapes whose topology is not known a priori. Geometric models are traditionally well suited for representing global shapes but not the local details. In this work, a powerful geometric shape meodeling scheme is investigated which allows for the representation of global shapes with local detail and permits model shaping via physics-based control as well as topological changes. These models are a blend of geometric and physics-based models, and are in spirit "similar" to the now popular deformable superquadrics, but differ from them considerably in their expressiveness and numerical stability, thereby promising greater applicability.

To understand evolving pathology in the central nervous system (CNS) and develop effective treatments, ways are needed to correlate the nerve fiber connectivity with the visualization of function. Such structure-function information is fundamental in CNS processes since anatomical connections determine where information is passed and processed. Recent methods of magnetic resonance diffusion tensor imaging (DTI) can provide the fundamental information required for viewing structural connectivity and can visualize fiber bundles in the brain in vivo. However, robust and accurate acquisition and processing algorithms are needed to accurately map the nerve connectivity. Automatic fiber tract mapping in the central nervous system (CNS) is a challenging problem for image processing since the data is noisy, making reliable estimation of the fiber tracts difficult. DTI data sets are large and present a formidable challenge in the design of efficient algorithms. In this proposal, we will develop novel, statistically robust and efficient algorithms for automatic fiber tract mapping in the CNS. The automatic fiber tract mapping problem will be solved in two phases, namely a data smoothing phase and a fiber tract mapping phase. In the former, smoothing will be achieved via a new nonlinear anisotropic diffusion algorithm which smooths the data while striving to retain all relevant detail. In the latter, a smooth 3D vector field indicating the dominant anisotropic direction at each spatial location is computed from the smoothed data. Fiber tracts will then be determined as the regularized integral curves of this vector field using efficient numerical methods. To validate the automatically estimated fiber tracts, we will establish the correlation between fiber tracts in fluorescence microscopy images of stained and excised rat spinal cord/brain and the estimated fiber tracts from the DTI data obtained in vivo. The utility of the method for pathology will then be tested on injured spinal cords and on previously acquired data sets of whole mouse, rat brains and isolated hearts.

DESCRIPTION (provided by applicant): This project aims to automatically recover and analyze the 3D hippocampal shape from human brain MRI and localize the epileptic focus to the appropriate temporal lobe. The hypothesis is that shape differences (not volume) between the left and right hippocampi will distinguish patients with epilepsy from healthy controls and identify the hemispheric location of the epileptic focus. We propose a three phase solution to test this hypothesis: (a) the development of a learning algorithm to create hippocampal shape and image atlases for use in segmentation, (b) automatic (atlas-based) hippocampal segmentation and validation, and (c) automatic classification of patient scans and validation of the classifier. The proposed new segmentation scheme will involve an atlas-based approach wherein the atlas is constructed from a prospective data set using a novel learning algorithm based on finding the atlas shape as the minimum distance fitted shape from the given population of fitted shapes. An MR image atlas learned similarly will be used to augment the learned shape prior. The learnt atlases will then be employed for estimating the non-rigid deformation field required to achieve an atlas-based segmentation of the unknown subject scan from an archive of retrospective data. A novel classifier based on the Kernel Fischer discriminant is proposed for automatically classifying subjects into groups corresponding to controls and those with epileptic foci localized to either the left or the right lobe. The segmentation and the classification algorithms will be validated on synthetic and real MR brain scans from an archive. The proposed algorithmic schemes have considerable potential for use in the segmentation and analysis of other anatomical structures and thus have utility for other disease states.

[unreadable] DESCRIPTION (provided by applicant): Epilepsy consists of more than 40 clinical syndromes affecting 50 million people worldwide. Approximately 25 to 30% of the patients receiving medication have inadequate seizure control. Progressive changes are suggested by the existence of a so-called silent interval, often years in duration, between CNS infection, head trauma, or prolonged seizure (status epilepticus) and the later appearance of epilepsy. This process is known as epileptogenesis and is thought of as a cascade of dynamic biological events altering the balance between excitation and inhibition in neural networks. Understanding these changes is key to preventing the onset of epilepsy. To this end, high angular resolution diffusion weighted MR-imaging (HARDI) offers the possibility to non-invasively track structural changes in limbic structures (dentate gyrus etc.). Our goal is to develop mathematical models and efficient algorithms to process HARDI data acquired from rat brains that have been imaged during the epileptogenetic period and derive structural signatures that can be used to predict the onset of epilepsy. Note that there is no precedence to this work on structural signatures for use in prediction of the onset of epilepsy. [unreadable] [unreadable] Our mathematical model characterizes multiple fiber tracts at a voxel by a continuous probability density over 2-tensors instead of the now popular multi-tensor model. In the absence of multiple fibers at a voxel, the proposed density model defaults to a Gaussian which characterizes the presence of a single fiber. The novelty of this formulation lies in relating the signal and the probability density of the 2-tensors via the well known Laplace transform and for the Wishart densities, leads to a closed form solution. Additionally we propose to segment the 3D lattice of probability densities to extract the ROI and map out the fibers which will be validated using histology data. Several novel properties (Renyi entropy etc.) constituting the structural signature characterizing the epileptogenetic period will then be computed from the segmented ROI. These features will then be used in a Kernel-based clustering to label clusters over the epileptogenetic period. Prediction will then be achieved for a novel data set via a Bayesian optimization scheme. Validation of the prediction results will be done on data for which onset times of epilepsy are already known. The proposed research will significantly advance our understanding of limbic system reorganization caused not only by prolonged seizures, but also the effects of recurrent seizures and further hippocampus damage. [unreadable] [unreadable] [unreadable]

Present day crops are the product of the intense selection applied by the first farming communities on the genetic variation of ancestral wild populations. This activity led to the accumulation of genes that control a group of traits known as the "domestication syndrome." These traits comprise visible changes in growth habit, organ size, seed dispersal, and environmental responses. However, with the exception of a few species that were domesticated for their edible roots, root traits have been neglected in domestication studies. Yet, morphological differences between roots of wild and cultivated forms exist and can be considered the product of indirect selection. This observation raises questions about both the number of genes responsible for these differences and their identity.

The main objective of this project is the identification of the genes controlling the root morphological changes associated with the domestication of Phaseolus vulgaris, the common bean, a New World species domesticated 8-10,000 years ago. The target genes and their chromosomal position will be identified through a statistical analysis of root traits (growth rates, branching, etc.) in progeny obtained from a cross between a wild and a domesticated bean. This analysis requires two pieces of information: a measurement of the root trait, and the genetic makeup of each individual in the progeny at previously mapped marker-genes covering all chromosomes. Root traits will be measured via 2D scans using a root scanner, and 3D scans using magnetic resonance imaging (MRI). Tools to evaluate 3D MRI root images quantitatively have been developed and will be expanded to create a computer generative 3D root model. The genetic makeup of the progeny will be obtained with the latest generation of high throughput sequencing and genotyping technologies. The genes controlling the domestication-associated root traits will be identified and located on a chromosome when statistical tests show that progeny with the wild version of the marker-gene exhibit significant differences in root traits from those with the domesticated version.

Identification of genes controlling root traits associated with domestication will have broader scientific and educational impacts. Knowledge about these genes will improve our understanding of the carbon balance of ecosystems in which the carbon storage in root mass is a major concern, and will also facilitate the genetic manipulation of roots for crop improvement. This project will produce a computer generative 3D root model, which will be a powerful teaching tool, and the starting point in the development of a gene-based root model. The training of undergraduate and graduate students from both the University of Florida and Florida A&M University, an institution with minorities who are under-represented in the scientific research community, will be one of the major components of this project.

The goal of this research is the investigation of a computational framework that allows for group-wise joint segmentation, smoothing and registration of images and shapes. With recent advances in sensor technology, images (shapes and other types of data) are being generated in abundance, and there is now a need for algorithms that operate and process images in collections instead of individually. In particular, this requires segmentation, smoothing and registration, three most important image processing operations, to be formulated in new ways that emphasize the relational aspects of their inputs. In addition, image data in computer vision applications are usually sampled from low-dimensional manifolds embedded in high-dimensional features spaces, and an important problem is to construct versatile and expressive computational models that exploit their geometries for solutions. The proposed computational framework addresses these two issues by formulating a variational framework that unifies smoothing, segmentation and registration. Specifically, it uses hypergraphs to model the multiple geometric relations among the inputs, and the three operations are integrated in one single discrete variational framework defined over a hypergraph. The proposed framework provides a foundation for several principled joint segmentation and registration algorithms for images and shapes that can guarantee crucial properties such as compatibility, consistency, unbiasedness and symmetry. Furthermore, it also provides a new and more discriminative numerical signature for 2D and 3D shapes that can be important for many shape-related vision applications such as shape recognition, shape retrieval and image-based medical diagnosis.

DESCRIPTION (provided by applicant): Establishing structure-function correlations is fundamental to understanding how information is processed in the central nervous system (CNS). Axonal connectivity is a key relationship that facilitates information transmission and reception within the CNS. Recently, diffusion weighted magnetic resonance imaging (DW-MRI) methods have been shown to provide fundamental information required for viewing structural connectivity and have allowed visualization of fiber bundles in the CNS in vivo. In this project, we propose to develop methods for extraction and analysis of these patterns from high angular resolution diffusion weighted images (HARDI) that is known to have better resolving power over diffusion tensor imaging (DTI). To this end, a biologically relevant and clinically important model has been chosen to study changes in the organization of fibers in the intact and injured spinal cord. Our hypothesis is that, changes in geometrical properties of the anatomical substrate, identifying the region of injury and neuroplastic changes in distant spinal segments, correlate with different magnitudes of injury and levels of locomotor recovery following spinal cord injury (SCI). Prior to hypothesis testing, we will denoise the HARDI data and then construct a normal atlas cord. Deformable registration and tensor morphometry between a normal atlas and an injured cord would be performed to provide a distinct signature for each type of behavior recovery associated with the SCI substrate. Validation of the hypothesis will be performed through systematic histological analysis of cord samples following acquisition of the HARDI data. Spinal cords will be cut and stained with fiber and cell stains to verify changes in anatomical organization that result from contusive injury (common in humans as well) to the spinal cord. A comparison between anatomical characteristics obtained from histological versus HARDI analysis will provide validation for the image analysis and the hypothesis. Three severities of spinal cord injuries will be produced (light, mild and moderate contusions) based upon normed injury device parameters. The structural signatures of these labeled data subsets will then be identified. Automatic classification of novel &injured cord HARDI data sets will then be achieved using a large margin classifier. Finally, HARDI data acquired over time will be analyzed in order to learn and predict the level of locomotor recovery by studying the structural changes over time and developing a dynamic model of structural transformations corresponding to each chosen class. We will use an auto-regressive model in the feature space to track and predict structural changes in SCI and correlate it to functional recovery. PUBLIC HEALTH RELEVANCE: This project involves the development of automated methods to extract morphological signatures that characterize changes in spinal cord injury (SCI) substrate estimated from Diffusion MRI scans of rats, and predict the functional recovery by correlating to behavioral studies. Although the various algorithms developed here are for analysis of SCI, they can be used in other applications such as traumatic brain injury, in tracking and predicting developmental changes etc. from diffusion MRI scans.

DESCRIPTION (provided by applicant): Establishing structure-function correlations is fundamental to understanding how information is processed in the central nervous system (CNS). Axonal connectivity is a key relationship that facilitates information transmission and reception within the CNS. Recently, diffusion weighted magnetic resonance imaging (DW-MRI) methods have been shown to provide fundamental information required for viewing structural connectivity and have allowed visualization of fiber bundles in the CNS in vivo. In this project, we propose to develop methods for extraction and analysis of these patterns from high angular resolution diffusion weighted images (HARDI) that is known to have better resolving power over diffusion tensor imaging (DTI). To this end, a biologically relevant and clinically important model has been chosen to study changes in the organization of fibers in the intact and injured spinal cord. Our hypothesis is that, changes in geometrical properties of the anatomical substrate, identifying the region of injury and neuroplastic changes in distant spinal segments, correlate with different magnitudes of injury and levels of locomotor recovery following spinal cord injury (SCI). Prior to hypothesis testing, we will denoise the HARDI data and then construct a normal atlas cord. Deformable registration and tensor morphometry between a normal atlas and an injured cord would be performed to provide a distinct signature for each type of behavior recovery associated with the SCI substrate. Validation of the hypothesis will be performed through systematic histological analysis of cord samples following acquisition of the HARDI data. Spinal cords will be cut and stained with fiber and cell stains to verify changes in anatomical organization that result from contusive injury (common in humans as well) to the spinal cord. A comparison between anatomical characteristics obtained from histological versus HARDI analysis will provide validation for the image analysis and the hypothesis. Three severities of spinal cord injuries will be produced (light, mild and moderate contusions) based upon normed injury device parameters. The structural signatures of these labeled data subsets will then be identified. Automatic classification of novel & injured cord HARDI data sets will then be achieved using a large margin classifier. Finally, HARDI data acquired over time will be analyzed in order to learn and predict the level of locomotor recovery by studying the structural changes over time and developing a dynamic model of structural transformations corresponding to each chosen class. We will use an auto-regressive model in the feature space to track and predict structural changes in SCI and correlate it to functional recovery.

This project develops efficient incremental algorithms are proposed for computing averages and other statistical quantities of interest from pools of data incrementally acquired. Many existing data acquisition and processing methods have reached a level of sophistication so as to be able to acquire and/or synthesize data that reside in curved spaces such as spheres, hyperboloids etc. As such data have become ubiquitous in many Science and Engineering fields, need for efficient statistical analysis of these data has emerged as an area of significant importance. Further, in this era of massive and continuous streaming data, samples of data are acquired sequentially over time. Hence, from an applications and computational efficiency perspective, the desired averaging algorithm ought to be amenable to incremental updates to accommodate the newly acquired data over time. The developed algorithms can be applied different applications, such as face recognition from videos, action recognition, trajectory averaging and clustering from videos, image and video restoration, pattern clustering and classification, etc. In the context of diagnostic medical imaging, methods developed in this project can be used to automatically discriminate between various disease classes, such as Parkinson's and Essential Tremor which are distinct types of movement disorders.

This research investigates a general framework for recursive computation of the intrinsic mean and the principal geodesic analysis on several commonly encountered manifolds such as the manifold of symmetric positive definite matrices, the Grassmann, the Stiefel manifolds, the hypersphere, the manifold of special orthogonal matrices, and several others. The research team applies the developed recursive framework of computing statistics from manifold-valued data to several tasks namely, atlas computation from diffusion MRI in Medical Imaging, inter-class discrimination between sub-types of a neuro-degenerative disorder using diffusion MRI, face and action recognition, image and shape retrieval in Computer Vision applications.

Magnetic resonance imaging (MRI) is the most widely used diagnostic imaging tool for detecting neurodegenerative disorders such as Parkinson's Disease. This project will develop new automated methods for detecting subtle effects that can be revealed by MRI, including changes in water diffusional properties of human brain tissue, and functional brain activity. To assess the deviation from the normal brains, a computationally efficient algorithm will be developed to construct a population-specific brain structural template from a normal brain population. Further, a new algorithm will be developed to facilitate the detection of Parkinson's using diffusion MRI data. Finally, novel algorithms for establishing the correlation between the information derived from diffusion and functional MRI data will be developed, enabling prediction of functional activity given the anatomical information and vice-versa. Inferring such a correlation will make it possible to predict functional changes due to changes in tissue microstructure caused by neurodegenerative disorders and vice-versa.

In summary, the precise project goals are: (i) To develop a computationally efficient template brain map construction algorithm for features derived from diffusion MRI. In this context, the ensemble average propagator (EAP), which captures both orientation and shape information of the diffusion process at each voxel in the diffusion MRI data, is proposed. Validation of the constructed template will be performed using standard evaluation metrics for template-based segmentation. (ii) To develop novel methods to automatically discriminate between control and Parkinson's groups using the EAP fields as well as Cauchy deformation tensors (that capture the changes in EAP fields). Validation of the classifier will be achieved using the standard leave-k-out strategy. (iii) To develop a novel algorithm for kernel-based nonlinear regression between EAP fields derived from diffusion MRI and scalar-valued fields derived from functional MRI activation maps. The algorithm will be able to predict the level of activation given the EAP fields and vice-versa. These predictions will be validated using a priori labeled data sets. Predicting functional responses from structural information and vice-versa will significantly impact treatment planning of patients with Parkinson's Disease and other neurodegenerative disorders. The multidisciplinary nature of this project will provide the opportunity to collectively train graduate students from diverse backgrounds in the STEM related fields of this project.