INVESTIGADOR(ES) PRINCIPAL(ES):

Hernán Felipe García Arias

0 horas

CODIGO CIE

E6-16-6

Mauricio Alexander Alvarez Lopez

Tutor

0 Horas

TIPO DE CONVOCATORIA

2015. Estudiantes De Posgrado

TIPO DE PROYECTO

Investigación Aplicada

OBJETIVO(S)

Objetivo General: -Desarrollar una Metodología para la estimación de grupos de correspondencias en formas 3D relacionadas con estructuras cerebrales utilizando modelos probabilísticos de variable latente Objetivos Específicos: - Desarrollar una metodología para caracterización de mallas poligonales que estimen descriptores de forma 3D basado en Scale Invariants Heat Kernels Signature - Desarrollar una metodología para estimar correspondencias representativas en formas 3D utilizando modelos Bayesianos no-paramétricos. - Contrastar la exactitud en la estimación de las correspondencias entre formas 3D utilizando análisis de regiones de Voronoi y metodologías comúnmente utilizadas en el estado del arte como registro deformable.

RESUMEN

The correspondence problem in neuroimaging analysis is a challenging research topic due to the importance to establish meaningful relations between any pair of brain structures (static registration problem) [35], or to analyze temporal changes of a given neurodegenerative disease (dynamic analysis of brain structures) [16]. However, similarity measures that can capture common information between objects are dicult to obtain [13]. The main reason of this problem, relies on the fact that brain structures are nonrigid objects that exhibit morphological changes among subjects (brain volumetry over a population) and shape deformations in presence of a given neurodegenerative disease (i.e. Alzheimer and Parkinson) [14]. Analyzing brain structures properties, such as shape volumetry or cortical thickness, is an important task in the study of Alzheimer's disease due to the importance of monitoring a set of structures (i.e. hippocampus, thalamus, and ventricles), analyze anatomic connectivity, and nding disease patterns over the brain cortex [58, 19, 9]. Nevertheless, the high variability of the brain patterns such as size, curvedness, and curvature, makes it necessary to compute the correspondences between objects in a group-wise manner [49]. Moreover, most of the correspondence methods for medical image problems are focused on computing dierent similarity metrics based on texture, being bag of words features [6], largest common point-sets [2] and geodesic contours [34, 9] the most representative methods in the state-of-art. However, these approaches work only in objects of the same size, which gives a poor accuracy in nonrigid matching processes [9]. In addition, a full relation of the matched features in all of their points must be computed (point-to-point correspondence) [8, 14]. In order to model the structure of a given object, methods for unsupervised object matching have been developed in the last years [30]. The aim of these methods is to establish meaningful correspondences in scenarios where shapes are described by non-rigids objects or the similarity measure between objects can not be computed [64]. Variational Bayesian matching [30] and Bayesian canonical correlation analysis [31] are some examples of these methods in which a given probabilistic framework is performed to model features between objects and establish the shape correspondences. Nonetheless, these methods only handles full correspondence frameworks (i.e. pointto- point matching) and linear analysis over the shape descriptors (i.e. appearance descriptors), which makes unsuitable to modeling shared information between non-rigid objects (i.e. tissue properties in MRI data) [59], and non-linearities of the represented medical data (i.e. shape descriptors such as Heat Kernel Signatures ) [7]. That is why this problem is still an open research topic in computer vision [14]. According to the above, it is clear that there are methodological problems related to the correspondence problem in the medical imaging eld. For this reason, there arises the following research question: >is it possible to develop a probabilistic method for shape correspondences analysis that allows temporal and structure learning of nonrigid shapes relevant in medical imaging problems?

ESTADO

Concluido

FECHA DE INICIO

18/01/2016

FECHA DE FINALIZACION

18/01/2018

PRODUCTOS

Bayesian Optimization for Fitting 3D Morphable Models of Brain Structures

Artículos en revista A1 ó A2

Nonlinear probabilistic latent variable models for groupwise Correspondence analysis in brain structures

Artículos en revista A1 ó A2

POSTER: Gaussian Process Dynamical Models for Multimodal Affect Recognition

Ponencia en evento especializado