Publicación: Clasificación y caracterización de manchas de Sustancia Blanca por Descriptores de Textura en imágenes de Resonancia Magnética, usando Redes Bayesianas y Máquinas de Vector Soporte
Cargando...
Fecha
2016-10-07
Editor/a
Director/a
Tutor/a
Coordinador/a
Prologuista
Revisor/a
Ilustrador/a
Derechos de acceso
Atribución-NoComercial-SinDerivadas 4.0 Internacional
info:eu-repo/semantics/openAccess
info:eu-repo/semantics/openAccess
Título de la revista
ISSN de la revista
Título del volumen
Editor
Universidad Nacional de Educación a Distancia (España). Escuela Técnica Superior de Ingeniería Informática. Departamento de Inteligencia Artificial.
Resumen
Actualmente no existe una evidencia plenamente clara sobre la relación de las WMHs con enfermedades neurológicas, lo que supone una cuestión de relevancia en investigación. Por este motivo, se considera muy importante detectar, cuantificar y caracterizar las WMHs en personas que sufren este tipo de patologías. La delineación manual de WMHs es una tarea con un alto coste en tiempo y propensa a errores humanos. Por tanto, es obvio y justificable buscar un método robusto y lo más automático posible, para reconocer este tipo de objeto amorfo. Este trabajo se basa en el resultado de una segmentación completamente automática llevada a cabo por el proyecto AMOS-2D, donde el objetivo es detectar y clasificar manchas de sustancia blanca a partir de imágenes de resonancia magnética, utilizando técnicas de aprendizaje automático, como una herramienta de ayuda al diagnóstico al experto humano. A partir de los objetos detectados por AMOS-2D se desarrolla un conjunto de librerías con el fin de reconocer WMHs por medio de descriptores espaciales, de forma y de textura, haciendo uso de algoritmos de aprendizaje automático como Redes Bayesianas y Máquinas de Vector Soporte. No obstante, el estudio se centra fundamentalmente en el análisis de discriminación de WMHs por descriptores de textura. Se concluye que el conjunto de datos de entrenamiento utilizado no permite resolver la clasificación con una tasa menor al 10% cuando se tienen en cuenta todas las instancias del mismo. Sin embargo, filtrando verticalmente el conjunto de datos, el rendimiento del clasificador mejora a costa de descompensar la proporción de valores de la clase. Como trabajo futuro, sería interesante hacer un estudio en 3D y comparar con los resultados obtenidos en este trabajo para poder valorar si el rendimiento del sistema mejora o no al añadir una dimensión espacial.
There is no currently a completely clear evidence on the relation between WMHs and neurological diseases, what implies a question of relevance in investigation. For this reason, it is very important to detect, quantify and characterize WMHs on people suffering this kind of pathologies. The manual WMH outlining is a time consuming task and prone to human mistakes. So, it is obvious and reasonable to search for a robust and as much automatic as possible method in order to recognize this kind of amorphous object. This work is based on the results from a fully automatic segmentation carried out by the project AMOS-2D, where the target is detecting and classifying white matter hyperintensities from magnetic resonance images, by using machine learning techniques, as a diagnostic support tool for the human expert. From the segmented objects by AMOS-2D, a set of libraries is developed with the aim of recognizing WMHs through spatial, shape and texture descriptors making use of machine learning algorithms as Bayesian Networks and Support Vector Machines. Nevertheless, the study is focused mainly on the discriminatory analysis of WMHs by texture characteristics. The conclusion is that the used training dataset does not allow us to solve the WMH classification with an error rate less than 10 percentage when all the dataset instances are taken into account. However, when applying some filters vertically to the dataset, the classifier performance improves at the expense of unbalancing the proportion of class values. As future work, a 3D study would be interesting and compare its results with the achieved ones of this work in order to appreciate whether the system performance improves or not when adding a spatial dimension.
There is no currently a completely clear evidence on the relation between WMHs and neurological diseases, what implies a question of relevance in investigation. For this reason, it is very important to detect, quantify and characterize WMHs on people suffering this kind of pathologies. The manual WMH outlining is a time consuming task and prone to human mistakes. So, it is obvious and reasonable to search for a robust and as much automatic as possible method in order to recognize this kind of amorphous object. This work is based on the results from a fully automatic segmentation carried out by the project AMOS-2D, where the target is detecting and classifying white matter hyperintensities from magnetic resonance images, by using machine learning techniques, as a diagnostic support tool for the human expert. From the segmented objects by AMOS-2D, a set of libraries is developed with the aim of recognizing WMHs through spatial, shape and texture descriptors making use of machine learning algorithms as Bayesian Networks and Support Vector Machines. Nevertheless, the study is focused mainly on the discriminatory analysis of WMHs by texture characteristics. The conclusion is that the used training dataset does not allow us to solve the WMH classification with an error rate less than 10 percentage when all the dataset instances are taken into account. However, when applying some filters vertically to the dataset, the classifier performance improves at the expense of unbalancing the proportion of class values. As future work, a 3D study would be interesting and compare its results with the achieved ones of this work in order to appreciate whether the system performance improves or not when adding a spatial dimension.
Descripción
Categorías UNESCO
Palabras clave
Citación
Centro
Facultades y escuelas::E.T.S. de Ingeniería Informática
Departamento
Inteligencia Artificial