Segmentación de la válvula pulmonar a partir de imágenes de tomografía cardiaca usando una estrategia basada en realce por similaridad local
No hay miniatura disponible
Fecha
2017
Autores
Vera, Miguel
Huérfano, Yoleidy
Contreras-Velásquez, Julio
Bermúdez, Valmore
Del Mar, Atilio
Cuberos, María
Vivas, Marisela
Bautista, Nahid
Saenz, Frank
Rodriguez, Jhoel
Título de la revista
ISSN de la revista
Título del volumen
Editor
Cooperativa servicios y suministros 212518 RS
Resumen
En el siguiente artículo se da a conocer el
uso de la estrategia similaridad local, en la
segmentación tridimensional (3D) de la válvula
pulmonar en 20 imágenes cardiacas de tomografía
computarizada multicapa, correspondientes al ciclo cardiaco
completo de un sujeto. La estrategia consta de las
siguientes etapas: a) pre-procesamiento, b) segmentación
y c) entonación de parámetros. La etapa a) se aplica, preliminarmente
al instante de diástole final y se divide en dos
fases denominadas: Filtrado y Definición de una región de
interés (ROI) y se emplea la técnica denominada realce por
similaridad local (LSE). La aplicación de estas fases tiene
por finalidad abordar los problemas de ruido, artefactos
y bajo contraste que poseen las mencionadas imágenes.
La etapa b) permite la segmentación de la válvula pulmonar,
mediante un algoritmo de agrupamiento denominado
crecimiento de regiones (RG) el cual es aplicado a las
imágenes pre-procesadas. El RG es inicializado con un vóxel
“semilla” el cual es detectado mediante un operador
de inteligencia artificial denominado máquinas de soporte
vectorial de mínimos cuadrados (LSSVM). Finalmente,
durante la etapa c), una métrica denominada coeficiente
de Dice (Dc) es utilizada para comparar las segmentaciones
obtenidas mediante la estrategia propuesta y la segmentación
generada, manualmente, por un cardiólogo.
La combinación de técnicas de filtrado que genera el Dc
más elevado considerando el instante de diástole se aplica
posteriormente a las 19 imágenes 3D restantes, obteniéndose
un Dc promedio comparable con el reportado en la
literatura especializada.
In the following article, the use of the local similarity strategy in the three-dimensional (3D) segmentation of the pulmonary valve in 20 cardiac multilayer computed tomography images corresponding to the complete cardiac cycle of a subject is reported. The strategy consists of the following stages: a) pre-processing, b) segmentation and c) intonation of parameters. Step a) is applied, preliminary to the final diastole instant and is divided into two phases called: Filtering and Definition of a region of interest (ROI) and using the so-called local similarity (LSE) technique. The application of these phases is intended to address the problems of noise, artifacts and low contrast that these images have. Stage b) allows segmentation of the pulmonary valve, using a clustering algorithm called region growth (RG) which is applied to the pre-processed images. The RG is initialized with a “seed” voxel which is detected by an artificial intelligence operator called least squares vector support machines (LSSVM). Finally, during step c), a metric called Dice coefficient (Dc) is used to compare the segmentations obtained by the proposed strategy and the segmentation generated manually by a cardiologist. The combination of filtering techniques that generates the highest Dc considering the instant of diastole is subsequently applied to the remaining 19 3D images, obtaining an average Dc comparable to that reported in the specialized literature.
In the following article, the use of the local similarity strategy in the three-dimensional (3D) segmentation of the pulmonary valve in 20 cardiac multilayer computed tomography images corresponding to the complete cardiac cycle of a subject is reported. The strategy consists of the following stages: a) pre-processing, b) segmentation and c) intonation of parameters. Step a) is applied, preliminary to the final diastole instant and is divided into two phases called: Filtering and Definition of a region of interest (ROI) and using the so-called local similarity (LSE) technique. The application of these phases is intended to address the problems of noise, artifacts and low contrast that these images have. Stage b) allows segmentation of the pulmonary valve, using a clustering algorithm called region growth (RG) which is applied to the pre-processed images. The RG is initialized with a “seed” voxel which is detected by an artificial intelligence operator called least squares vector support machines (LSSVM). Finally, during step c), a metric called Dice coefficient (Dc) is used to compare the segmentations obtained by the proposed strategy and the segmentation generated manually by a cardiologist. The combination of filtering techniques that generates the highest Dc considering the instant of diastole is subsequently applied to the remaining 19 3D images, obtaining an average Dc comparable to that reported in the specialized literature.
Descripción
Palabras clave
Válvula pulmonar, Procesos de filtrado, Segmentación, Realce por similaridad local, Pulmonary valve, Filtering processes, Segmentation, Local similarity enhancement