Publicación: Aplicación de técnicas de aprendizaje profundo a la segmentación de imagen ecocardiográfica
Cargando...
Fecha
2023-09
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 Técnica Superior de Ingeniería Informática. Departamento de Inteligencia Artificial
Resumen
La fiebre reumática es una enfermedad infecciosa que afecta con mayor frecuencia a los niños de 5 a 15 años, aunque se puede presentar también en adultos y muy rara vez en niños más pequeños. Suele causar daños en varias partes del cuerpo, especialmente en las válvulas del corazón, dando lugar así a la enfermedad cardíaca reumática (ECR). Se estima que en la actualidad esta enfermedad afecta a más de 30 millones de personas en el mundo y causa más de 300.000 muertes anuales, muchas de ellas de personas menores de 25 años, además de provocar invalidez permanente en muchos casos. Los problemas que ocasiona son más frecuentes durante el embarazo y el parto. En países endémicos la prevención primaria es complicada principalmente por la falta de acceso a servicios médicos cualificados por lo que la prevención secundaria, consistente en realizar un cribado mediante ecocardiografía para detectar los pacientes asintomáticos y aplicarles tratamiento profiláctico con penicilina, ha demostrado ser una vía de acción más efectiva. El principal objetivo de esta línea de TFM es diseñar un sistema inteligente que, mediante técnicas de inteligencia artificial aplicadas al procesamiento de imágenes ecocardiográficas, sirva de ayuda al diagnóstico de la enfermedad cardíaca reumática. He comparado cuatro arquitecturas diferentes: una red neuronal convolucional (CNN) con tres capas convolucionales, una ResUnet, U-Net y Laddernet. He probado diferentes técnicas para mejorar el rendimiento de los modelos y reducir sus problemas de generalización, incluyendo preprocesar las imágenes con filtro CLAHE, filtro ecualizado, filtro gaussiano, filtro de Sobel y aumento de datos. Así mismo he hecho pruebas de todas las arquitecturas con diferentes tamaños de kérnel 2, 3 y 4. Finalmente, en el conjunto de pruebas, he obtenido un valor de coeficiente DICE 0,805 para la arquitectura CNN propuesta; 0,890 para la arquitectura Laddernet, 0,895 para ResUnet, 0,894 para una red U-Net preentrenada y 0,911 para U-Net con aumento de datos.
Rheumatic fever is an infectious disease that most commonly affects children aged 5 to 15 years, although it can also occur in adults and very rarely in younger children. It usually causes damage to various parts of the body, especially the heart valves, thus giving rise to rheumatic heart disease (RHD). It is estimated that this disease currently affects more than 30 million people worldwide and causes more than 300,000 deaths annually, many of them in people under 25 years of age, as well as causing permanent disability in many cases. The problems it causes are most frequent during pregnancy and childbirth. In endemic countries, primary prevention is complicated mainly by the lack of access to qualified medical services, so secondary prevention, consisting of screening by echocardiography to detect asymptomatic patients and apply prophylactic treatment with penicillin, has proven to be a more effective course of action. The main objective of this line of TFM is to design an intelligent system that, by means of artificial intelligence techniques applied to echocardiographic image processing, will aid in the diagnosis of rheumatic heart disease. I have compared four different architectures: an convolutional neural network (CNN) with three convolutional layers, a ResUnet, U-Net and Laddernet. I have tested different techniques to improve the performance of the models and reduce their generalization problems, including preprocessing the images with CLAHE filter, equalized filter, Gaussian filter, Sobel filter and data augmentation. Likewise, I have tested all architectures with different kérnel sizes 2, 3 and 4. Finally, in the test dataset, I obtained a DICE coefficient value 0.805 for the proposed CNN architecture; 0.890 for the Laddernet architecture, 0.895 for ResUnet, 0.894 for a pre-trained UNet and 0.911 for U-Net with data augmentation.
Rheumatic fever is an infectious disease that most commonly affects children aged 5 to 15 years, although it can also occur in adults and very rarely in younger children. It usually causes damage to various parts of the body, especially the heart valves, thus giving rise to rheumatic heart disease (RHD). It is estimated that this disease currently affects more than 30 million people worldwide and causes more than 300,000 deaths annually, many of them in people under 25 years of age, as well as causing permanent disability in many cases. The problems it causes are most frequent during pregnancy and childbirth. In endemic countries, primary prevention is complicated mainly by the lack of access to qualified medical services, so secondary prevention, consisting of screening by echocardiography to detect asymptomatic patients and apply prophylactic treatment with penicillin, has proven to be a more effective course of action. The main objective of this line of TFM is to design an intelligent system that, by means of artificial intelligence techniques applied to echocardiographic image processing, will aid in the diagnosis of rheumatic heart disease. I have compared four different architectures: an convolutional neural network (CNN) with three convolutional layers, a ResUnet, U-Net and Laddernet. I have tested different techniques to improve the performance of the models and reduce their generalization problems, including preprocessing the images with CLAHE filter, equalized filter, Gaussian filter, Sobel filter and data augmentation. Likewise, I have tested all architectures with different kérnel sizes 2, 3 and 4. Finally, in the test dataset, I obtained a DICE coefficient value 0.805 for the proposed CNN architecture; 0.890 for the Laddernet architecture, 0.895 for ResUnet, 0.894 for a pre-trained UNet and 0.911 for U-Net with data augmentation.
Descripción
Categorías UNESCO
Palabras clave
Citación
Centro
Facultades y escuelas::E.T.S. de Ingeniería Informática
Departamento
Inteligencia Artificial