Publicación: Segmentación automática de la red de vasos sanguíneos en imágenes de retina mediante redes neuronales artificiales evolutivas y operadores basados en “Local Binary Patterns” (LBP)
Cargando...
Fecha
2013-09-01
Autores
Editor/a
Director/a
Tutor/a
Coordinador/a
Prologuista
Revisor/a
Ilustrador/a
Derechos de acceso
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
En este trabajo de Fin de Máster se presenta un nuevo método para la segmentación de los vasos sanguíneos de la retina en fotografías de fondo de ojo, basado en la utilización de Redes Neuronales Artificiales (RNA) y en el uso del operador LBP (Local Binnary Pattern), un operador que, tradicionalmente, ha sido muy utilizado en el análisis de imágenes para la detección de texturas. En realidad, para ser más precisos, se ha evaluado no sólo el operador LBP clásico sino, además, un conjunto de operadores que son variaciones del original. Algunas de estas variaciones corresponden a operadores ya existentes en la literatura y otras han sido aquí propuestas. Aplicando este conjunto de operadores a las fotografías de fondo de ojo se obtiene, para cada pixel de las mismas, un conjunto de valores que conforman un vector de características. A partir de vectores de este tipo, se construirá un conjunto de datos para entrenar una RNA. Sin embargo, no se utilizará un método clásico de entrenamiento de RNAs sino que el aprendizaje de la RNA óptima se hará a partir de un algoritmo evolutivo basado en Grammatical Evolution. Mediante este paradigma evolutivo, no sólo se aprenderá la RNA más adecuada para clasificar cada pixel de la imagen como perteneciente o no a un vaso sanguíneo, sino que se seleccionará además, de forma automática, la topología de la red y las características de entrada más adecuadas para ello, simplificando así la tarea de determinar cuáles son las variantes de LBP más discriminantes que han de ser utilizadas como entrada a la red. La evaluación del método se ha llevado a cabo con dos de las bases de datos de imágenes digitales de retina más utilizadas al abordar la solución de este tipo de problema: DRIVE (Digital Retinal Images for Vessel Extraction), y STARE (STructured Analysis of the Retina). El método obtiene unos resultados competitivos respecto a otros métodos existentes en la literatura, tanto desde el punto de vista de precisión y sensibilidad como desde el punto de vista del coste computacional.
In this Master’s Thesis a new method is presented for automated segmentation of retinal blood vessels in fundus photographs based on the use of Artificial Neural Networks (ANN) and in the use of LBP operator (Local Binary Patterns), an operator that has traditionally been used in image analysis for the detection of textures. Actually, to be precise, it has been evaluated not only the classic LBP operator but also a set of operators which are variations of the original. Some of these variations correspond to existing operators in the literature and others have been proposed here. Applying this set of operators to the fundus photographs is obtained, for each pixel of the same, a set of values that form a feature vector. From such vectors, will be build a data set to train an ANN. However, don't will be used a classical method for training ANN, but learning the optimal ANN will be made from an evolutionary algorithm based on Grammatical Evolution. Through this evolutionary paradigm, will be not only learned the most appropriate ANN to classify each image pixel as belonging or not to a blood vessel, but also selected, automatically, the network topology and the input characteristics more suitable for this purpose, thus simplifying the task of determining the most discriminant LBP variants to be used as input to the network. The evaluation of the method was carried out with two of most used digital retinal images databases to approach the solution of this type of problem: DRIVE (Digital Retinal Images for Vessel Extraction), and STARE (Structured Analysis of the Retina). The method obtains competitive performance over other methods available in the literature, both in terms of accuracy and sensitivity, and from the standpoint of computational cost.
In this Master’s Thesis a new method is presented for automated segmentation of retinal blood vessels in fundus photographs based on the use of Artificial Neural Networks (ANN) and in the use of LBP operator (Local Binary Patterns), an operator that has traditionally been used in image analysis for the detection of textures. Actually, to be precise, it has been evaluated not only the classic LBP operator but also a set of operators which are variations of the original. Some of these variations correspond to existing operators in the literature and others have been proposed here. Applying this set of operators to the fundus photographs is obtained, for each pixel of the same, a set of values that form a feature vector. From such vectors, will be build a data set to train an ANN. However, don't will be used a classical method for training ANN, but learning the optimal ANN will be made from an evolutionary algorithm based on Grammatical Evolution. Through this evolutionary paradigm, will be not only learned the most appropriate ANN to classify each image pixel as belonging or not to a blood vessel, but also selected, automatically, the network topology and the input characteristics more suitable for this purpose, thus simplifying the task of determining the most discriminant LBP variants to be used as input to the network. The evaluation of the method was carried out with two of most used digital retinal images databases to approach the solution of this type of problem: DRIVE (Digital Retinal Images for Vessel Extraction), and STARE (Structured Analysis of the Retina). The method obtains competitive performance over other methods available in the literature, both in terms of accuracy and sensitivity, and from the standpoint of computational cost.
Descripción
Categorías UNESCO
Palabras clave
segmentación de vasos sanguíneos, imágenes de retina, redes neuronales artificiales, Local Binary Patterns, grammatical evolution, segmentation of blood vessels, retinal images, artificial neural networks
Citación
Centro
Facultades y escuelas::E.T.S. de Ingeniería Informática
Departamento
Inteligencia Artificial