Carotid artery ultrasound image-based cardiovascular risk prediction using deep learning
- Autores: Maria del Mar Vila Muñoz
- Directores de la Tesis: Laura Igual Muñoz (dir. tes.), María Grau Magaña (codir. tes.), Petia Radeva Ivanova (tut. tes.)
- Lectura: En la Universitat de Barcelona ( España ) en 2024
- Idioma: español
- Tribunal Calificador de la Tesis: Debora Gil Resina (presid.), Oliver Diaz (secret.), Marcelino Bermúdez López (voc.)
- Programa de doctorado: Programa de Doctorado en Matemáticas e Informática por la Universidad de Barcelona
- Materias:
- Enlaces
- Tesis en acceso abierto en: TDX
- Resumen
La tesis aborda las enfermedades cardiovasculares, en particular la aterosclerosis, usando imágenes de ultrasonido de la arteria carótida.
Introduce métodos de Aprendizaje Profundo para mejorar la caracterización de las placas ateroscleróticas y la predicción del riesgo cardiovascular.
Además, propone un modelo de supervivencia que utiliza características de imágenes para reclasificar a individuos de moderado a alto riesgo.
Cardiovascular Diseases (CVDs), the leading cause of death in developed countries, often involve atherosclerosis, which is a chronic inflammatory thickening of the inner artery layer. Monitoring atherosclerotic plaque detection and its characteristics is crucial for assessing future cardiovascular events.
Carotid Artery (CA) Ultrasound (US) images are utilized for subclinical atherosclerosis detection by measuring carotid Intima Media Thickness (IMT) and identifying atherosclerotic plaques. This thesis introduces Deep Learning (DL) methods to segment CA US images and characterize atherosclerotic plaque, aiming to improve cardiovascular risk prediction.
First, we address the segmentation of the Carotid Intima-Media (CIM) region, where the IMT is estimated. In this work, we introduce a fully automated method based on Convolutional Neural Networks that accurately localizes the carotid IMT region in longitudinal B-mode CA US images. In particular, we present a novel single-step approach using DenseNets for semantic segmentation, resulting in enhanced subclinical atherosclerosis detection through efficient carotid IMT estimation and atherosclerotic plaque detection.
This thesis introduces two clinical applications of carotid IMT estimation and atherosclerotic plaque detection. The first study evaluates cardiovascular event risk in autoimmune disease patients, focusing on chronic inflammation's impact on subclinical atherosclerosis. In the second study, we examine the coexistence of subclinical atherosclerosis in the lower limb (Ankle-Brachial Index) and carotid arteries. The findings of both studies highlights the systemic nature of atherosclerosis, suggesting a correlation between biomarkers in different areas and the likelihood of subclinical disease.
Finally, we explore new ways of improving cardiovascular risk prediction using DL techniques to extract information from CA US. In cardiovascular epidemiology, risk prediction functions assess the likelihood of a cardiovascular event based on individual clinical variables, using survival models. Despite their accurate stratification into low, moderate, and high-risk groups, a significant number of cardiovascular events still occur in the medium-risk category. This study introduces a novel approach for CA characterization, integrating individual artery condition data into traditional survival models. The work presents an innovative survival model that incorporates CA US image features derived from Deep Neural Networks, enabling effective cardiovascular risk prediction and the reclassification of individuals from the moderate to the high-risk category within the survival model.
