Publicación: Detección y segmentación simultánea de estructuras anatómicas en imágenes de retina basada en modelos de conocimiento relacionales Intra e Inter-Estructura
Cargando...
Fecha
2018-07-17
Autores
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 Internacional de Doctorado. Programa de Doctorado en Sistemas Inteligentes
Resumen
La detección y segmentación de estructuras anatómicas en imágenes de retina constituyen dos tareas previas de gran importancia para la implementación de cualquier sistema automático de diagnóstico de patologías retinianas, siendo hoy día dos problemas aún abiertos. Por ejemplo, los resultados de detección de las diferentes estructuras anatómicas pueden utilizarse como un paso previo a la segmentación o como puntos de referencia usados por sistemas automáticos para el diagnóstico de diferentes patologías retinianas. Por otra parte, la segmentación del disco óptico es fundamental, por ejemplo, para el diagnóstico de enfermedades como el glaucoma o la retinopatía diabética proliferativa. Igualmente, la segmentación de la fóvea permite, por ejemplo, graduar la gravedad de determinadas maculopatías, permitiendo calcular la distancia de determinados patrones de patología al centro de esta estructura. La mayoría de los trabajos en la literatura relacionada han estado orientados a la detección/segmentación de cada estructura individualizada, o a la detección/segmentación de dos o más estructuras realizada de forma secuencial. Por el contrario, en esta tesis, se proponen dos metodologías, una de detección y otra de segmentación, que comparten tres propiedades importantes: (i) son independientes del tamaño de la imagen de entrada; (ii) están orientadas al uso conjunto del conocimiento relacional asociado a cada estructura retiniana (CRintra-e) y al de varias de estas estructuras (CRinter-e); y (iii) el uso conjunto de ambos tipos de conocimiento permite realizar la detección/segmentación de n estructuras de forma simultánea. La primera característica persigue que la configuración de parámetros de cualquier método basado en estas metodologías sea independiente del tamaño de imagen y, la segunda y tercera, que el número de falsos positivos disminuya respecto de aquellos casos en que la detección o segmentación se realiza de forma individual o secuencial. A partir de las dos metodologías mencionadas anteriormente, se detallan los pasos seguidos para la implementación de sendos métodos de detección y segmentación de estructuras retinianas. Uno de estos pasos hace referencia a la construcción de una serie de modelos relacionales de cada una de las estructuras a detectar/segmentar, distinguiendo entre modelos que solo usan información de la propia estructura (modelos intra-e) o de más de una estructura (modelos inter-e). De esta forma, se construyen cuatro modelos intra-e (disco óptico, mácula, red de vasos principales y haz vascular) y un modelo implícito inter-e (basado en restricciones) que son usados por el primer método para la detección simultánea de las cuatro estructuras mencionadas. De otro lado, se construyen dos modelos intra-e (disco óptico y fóvea) y un modelo relacional para ser usado por el segundo método para la segmentación simultánea de las dos estructuras mencionadas. La principal diferencia entre los modelos intra-e usados para detección y segmentación radica en su complejidad, siendo muy simples en el primer caso y algo más complejos en el segundo. Los resultados de los experimentos diseñados refrendan las bondades de la simultaneidad y el uso conjunto del CRintra e inter-e, tanto en detección como en segmentación, siendo la disminución de la tasa de falsos positivos una de sus principales consecuencias. Todos los experimentos se realizaron en bases de imágenes públicas para facilitar la comparación con otros métodos de detección/segmentación. Un experimento adicional muestra que la incorporación del resultado de detección como entrada adicional al método de segmentación favorece los resultados de este último, a costa de un pequeño incremento del tiempo empleado en el proceso.
Detection and segmentation of anatomical structures in retinal images are two previously required tasks for the implementation of any automatic system oriented to the diagnosis of retinal pathologies. Despite the fact of their outstanding importance, both tasks remain open problems to this day. As an example, results of the detection for the dierent anatomic structures can be used as a prior step to segmentation, or even as landmarks for automatized diagnosis systems with the purpose of detecting any illness symptomatology. On the other hand, optic disk segmentation is of capital importance to detect diseases as glaucoma or proliferative diabetic retinopathy. Equally, fovea segmentation enables the measurement of the distance to possible disease patterns, evaluating the grade of certain maculopathies. Most of the literary works related to this subject have been developed in two diferent ways: focusing in the detection/segmentation of each structure as an individualized item or carrying out the detection/segmentation of two or more structures in a sequential way. Conversely, this thesis proposes two methodologies, one for detection and another for segmentation, sharing both of them three relevant features: (i) they are independent of the size of the input image; (ii) they are oriented to the joint use of the relational knowledge associated with each retinal structure (intra-SRK) and with other structures (inter-SRK); and (iii) the joint use of both types of knowledge allows the execution of the detection/segmentation of n structures simultaneously. The rst characteristic intends the parameter conguration of any method based in these methodologies to be independent from the image size. In addition, the second and third features are aimed to diminish the number of false positives in relation to those cases in which just one structure is detected/segmented or detection/segmentation iii is sequentially implemented. On the basis of the two above mentioned methodologies, the necessary steps to implement a method for retinal structures detection/segmentation are described. One of these steps is associated with the construction of relational models for each one of the structures destined to be detected/segmented, distinguishing between models that employ knowledge from just one structure (intra-SRK models) and those that use knowledge from more than one structure (inter-SRK model). This way, four intra-SRK models are built (being related to optic disk, macula, main retinal vessel network and vascular bundle), plus an implicit inter-SRK model (based on restrictions). All of them are used by the rst method to simultaneously detect the four above mentioned structures. On the other side, two intra-SRK models (optic disk and fovea) and a relational model are used by the second method for simultaneously segment the two structures previously referred. The main dierence between intra-SRK models used for detection and segmentation lies in its degree of complexity, resulting in relatively simple models in the rst case, but somewhat more complex in the second. The results from the designed experiments show the benets of simultaneity property and the joint use of intra-SRK and inter-SRK for both detection and segmentation, being the decrease of the false positive rate its main consequence. All experiments were carried out on public databases to make easier the comparison to other detection/segmentation methods. An additional experiment shows that the incorporation of the detection result as an additional input for the segmentation method favors the results obtained by this last one, at the expense of a little increase in the computational time spent to complete the process.
Detection and segmentation of anatomical structures in retinal images are two previously required tasks for the implementation of any automatic system oriented to the diagnosis of retinal pathologies. Despite the fact of their outstanding importance, both tasks remain open problems to this day. As an example, results of the detection for the dierent anatomic structures can be used as a prior step to segmentation, or even as landmarks for automatized diagnosis systems with the purpose of detecting any illness symptomatology. On the other hand, optic disk segmentation is of capital importance to detect diseases as glaucoma or proliferative diabetic retinopathy. Equally, fovea segmentation enables the measurement of the distance to possible disease patterns, evaluating the grade of certain maculopathies. Most of the literary works related to this subject have been developed in two diferent ways: focusing in the detection/segmentation of each structure as an individualized item or carrying out the detection/segmentation of two or more structures in a sequential way. Conversely, this thesis proposes two methodologies, one for detection and another for segmentation, sharing both of them three relevant features: (i) they are independent of the size of the input image; (ii) they are oriented to the joint use of the relational knowledge associated with each retinal structure (intra-SRK) and with other structures (inter-SRK); and (iii) the joint use of both types of knowledge allows the execution of the detection/segmentation of n structures simultaneously. The rst characteristic intends the parameter conguration of any method based in these methodologies to be independent from the image size. In addition, the second and third features are aimed to diminish the number of false positives in relation to those cases in which just one structure is detected/segmented or detection/segmentation iii is sequentially implemented. On the basis of the two above mentioned methodologies, the necessary steps to implement a method for retinal structures detection/segmentation are described. One of these steps is associated with the construction of relational models for each one of the structures destined to be detected/segmented, distinguishing between models that employ knowledge from just one structure (intra-SRK models) and those that use knowledge from more than one structure (inter-SRK model). This way, four intra-SRK models are built (being related to optic disk, macula, main retinal vessel network and vascular bundle), plus an implicit inter-SRK model (based on restrictions). All of them are used by the rst method to simultaneously detect the four above mentioned structures. On the other side, two intra-SRK models (optic disk and fovea) and a relational model are used by the second method for simultaneously segment the two structures previously referred. The main dierence between intra-SRK models used for detection and segmentation lies in its degree of complexity, resulting in relatively simple models in the rst case, but somewhat more complex in the second. The results from the designed experiments show the benets of simultaneity property and the joint use of intra-SRK and inter-SRK for both detection and segmentation, being the decrease of the false positive rate its main consequence. All experiments were carried out on public databases to make easier the comparison to other detection/segmentation methods. An additional experiment shows that the incorporation of the detection result as an additional input for the segmentation method favors the results obtained by this last one, at the expense of a little increase in the computational time spent to complete the process.
Descripción
Categorías UNESCO
Palabras clave
Citación
Centro
Facultades y escuelas::E.T.S. de Ingeniería Informática
Departamento
Grupo de investigación
Grupo de innovación
Programa de doctorado
Programa de doctorado en sistemas inteligentes