Fecha
2025-06
Editor/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 (UNED). E.T.S. de Ingeniería Informática
Citas
0 citas en 
0 citas en 
Resumen
La enfermedad cardíaca reumática es una complicación grave de la fiebre reumática que causa daños en las válvulas del corazón. La detección y tratamiento tempranos de esta enfermedad pueden reducir significativamente la mortalidad asociada a ella. Es importante utilizar las nuevas tecnologías, como la inteligencia artificial (IA), para poder abordar estos problemas en países donde esta enfermedad es endémica y se carece de personal cualificado para su detección. La correcta interpretación de ecocardiografías para su diagnóstico de forma automatizada, para realizar una previa clasificación y reducir el enorme trabajo que supone para los profesionales, es un paso importante en este camino. Este Trabajo de fin de máster forma parte del proyecto CES-BC-RHD, cuyo objetivo es desarrollar herramientas basada en aprendizaje profundo para ayudar al diagnóstico de esta enfermedad en países endémicos. En él, abordamos la evaluación de distintas arquitecturas y técnicas de optimización de modelos de aprendizaje profundo para la categorización de cada fotograma de una ecocardiografía en las distintas fases del ciclo cardíaco. Esto, mediante técnicas de integración con otros modelos de aprendizaje automático, podría ayudar al diagnóstico de la enfermedad cardíaca reumática. Es justo en el cambio de sístole a diástole (y viceversa) cuando más información relevante arrojan los ecocardiogramas.
Se han estudiado tres arquitecturas distintas: una 3D-CNN, y dos CNN+RNN; y se utilizó Optuna para la optimización de hiperparámetros. Los modelos se entrenaron con el conjunto de datos públicos TED, con 98 pacientes. Se evaluaron distintas técnicas de mejora como el uso de filtro CLAHE, aumento de datos o el flujo óptico, entre otras. Para esta evaluación se utilizó validación cruzada sobre un conjunto de datos de entrenamiento, dejando 10 pacientes para la fase de test. Se evaluó la generalización de los modelos en 200 vídeos de EchoNet-Dynamic, donde se estimó el ciclo completo mediante una heurística basada en el índice de similitud estructural (SSIM). El flujo óptico, técnica que analiza el movimiento de objetos en una secuencia de imágenes, ha demostrado ser el componente con mayor impacto en los modelos, llegando a incrementar en más de un 18 % los resultados obtenidos en términos de F1-score. La mejor configuración alcanzó un F1-score de 0.95, Accuracy de 0.95 y AUC de 0.98 en nuestro conjunto de test. Aunque fue otro modelo el que mostró mejor generalización en EchoNet, con un F1-score de 0.93 y AUC de 0.98. El aumento de datos redujo la varianza entre folds.
La mayor limitación ha sido la falta de un conjunto de datos grande, ya que el que hemos usado tiene apenas un centenar de pacientes y un sólo ciclo cardíaco completo por paciente. En líneas futuras podrían explorarse la codificación temporal utilizando transformadores, distintas técnicas de fusión de flujo óptico y explorar la ayuda al etiquetado utilizando técnicas como SSIM.
Rheumatic heart disease is a serious complication of rheumatic fever that causes damage to the heart valves. Early detection and treatment of this disease can significantly reduce the associated mortality. Leveraging new technologies, such as artificial intelligence (AI), is therefore crucial in countries where the disease is endemic and qualified personnel for its diagnosis are scarce. Automated, reliable interpretation of echocardiograms to pre-classify studies and lighten the workload of clinicians is a key step toward this goal. This Master’s Thesis is part of the CES-BC-RHD project, whose aim is to develop deep learning–based tools to support the diagnosis of this disease in endemic regions. In this work, we address the evaluation of different architectures and optimization techniques for deep learning models aimed at classifying each frame of an echocardiogram into the phases of the cardiac cycle. Combined with other machine learning models, this cate- gorisation could facilitate the rheumatic heart disease diagnosis, as the transitions between systole and diastole (and vice versa) reveal the most clinically relevant information. Three different architectures were studied: a 3D-CNN and two CNN+RNN. Optuna was used for hyperparameter optimization. Models were trained on the public TED dataset comprising 98 patients. Several enhancement techniques were assessed, CLAHE contrast filtering, data augmentation, optical flow, among others. Performance was measured with cross-validation on the training set, reserving 10 patients for a hold-out test set. Generalization was evaluated on 200 videos from EchoNet-Dynamic, where the complete cycle was estimated using a heuristic based on the Structural Similarity Index (SSIM). Optical flow, a technique that analyzes the movement of objects in a sequence of images, has proven to be the component with the greatest impact on the models, increasing the results by more than 18% in terms of F1-score. The best configuration achieved an F1-score of 0.95, accuracy of 0.95, and AUC of 0.98 on our test set. However, a different model showed better generaliza- tion on EchoNet, achieving an F1-score of 0.93 and AUC of 0.98. Data augmentation mainly reduced fold-to-fold variance. The main limitation was the small dataset compising fewer than one hundred patients and only a single cardiac cycle per patient. Future work could explore transformer-based temporal encoders, advanced optical flow fusion strategies, and semi-automated labelling assitance using metrics such as SSIM.
Rheumatic heart disease is a serious complication of rheumatic fever that causes damage to the heart valves. Early detection and treatment of this disease can significantly reduce the associated mortality. Leveraging new technologies, such as artificial intelligence (AI), is therefore crucial in countries where the disease is endemic and qualified personnel for its diagnosis are scarce. Automated, reliable interpretation of echocardiograms to pre-classify studies and lighten the workload of clinicians is a key step toward this goal. This Master’s Thesis is part of the CES-BC-RHD project, whose aim is to develop deep learning–based tools to support the diagnosis of this disease in endemic regions. In this work, we address the evaluation of different architectures and optimization techniques for deep learning models aimed at classifying each frame of an echocardiogram into the phases of the cardiac cycle. Combined with other machine learning models, this cate- gorisation could facilitate the rheumatic heart disease diagnosis, as the transitions between systole and diastole (and vice versa) reveal the most clinically relevant information. Three different architectures were studied: a 3D-CNN and two CNN+RNN. Optuna was used for hyperparameter optimization. Models were trained on the public TED dataset comprising 98 patients. Several enhancement techniques were assessed, CLAHE contrast filtering, data augmentation, optical flow, among others. Performance was measured with cross-validation on the training set, reserving 10 patients for a hold-out test set. Generalization was evaluated on 200 videos from EchoNet-Dynamic, where the complete cycle was estimated using a heuristic based on the Structural Similarity Index (SSIM). Optical flow, a technique that analyzes the movement of objects in a sequence of images, has proven to be the component with the greatest impact on the models, increasing the results by more than 18% in terms of F1-score. The best configuration achieved an F1-score of 0.95, accuracy of 0.95, and AUC of 0.98 on our test set. However, a different model showed better generaliza- tion on EchoNet, achieving an F1-score of 0.93 and AUC of 0.98. Data augmentation mainly reduced fold-to-fold variance. The main limitation was the small dataset compising fewer than one hundred patients and only a single cardiac cycle per patient. Future work could explore transformer-based temporal encoders, advanced optical flow fusion strategies, and semi-automated labelling assitance using metrics such as SSIM.
Descripción
Categorías UNESCO
Palabras clave
Ecocardiograma, Aprendizaje Profundo, Clasificación de fases cardíacas, Diagnóstico asistido por IA, Flujo óptico, Redes 3D-CNN, Redes CNN-RNN, Echocardiogram, Deep Learning, Cardiac Phase Classification, AI-assisted Diagnosis, Optical Flow, 3D-CNN Networks, CNN-RNN Networks
Citación
Román Garzón, Juan Alonso. Trabajo Fin de Máster: "Estudio de arquitecturas espacio-temporales de aprendizaje profundo para la detección de fase cardíaca". Universidad Nacional de Educación a Distancia (UNED) 2025
Centro
E.T.S. de Ingeniería Informática
Departamento
Inteligencia Artificial