Publicación: Characterization of left ventricle sarcomere properties using deep learning
Cargando...
Fecha
2023
Autores
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 (España). Facultad de Ciencias
Resumen
Independientemente de su origen, la disfunción diastólica está presente en la práctica totalidad de las enfermedades estructurales del miocardio. Por ello, es de especial relevancia para la práctica clínica el conocer los cambios que pueden producirse en las propiedades mecánicas del corazón en estas circunstancias, especialmente en el ventrículo izquierdo, más propenso al fallo al trabajar a mayor presión. Actualmente, existen modelos directos que relacionan las propiedades del sarcómero con las propiedades de las cámaras y su función. No obstante, dado que las propiedades del sarcómero son prácticamente imposibles de obtener in vivo o in silico, no existen modelos que solucionen el problema de calibración inverso. La presente tesina tiene como objetivo evaluar diferentes redes neuronales convolucionales que predigan las propiedades del sarcómero en base a pares de variables hemodinámicas sintéticas, obtenidas mediante CircAdapt, en una metodología de transfer learning. Los resultados de estos análisis muestran que, cuando las señales no incorporan ruido blanco, los niveles de exactitud son muy elevados, alrededor del 90%, para cualquiera de las parejas de variables analizadas. Al introducir ruido blanco en las señales, la combinación de presión y volumen del ventrículo izquierdo mantiene este elevado nivel de exactitud. La precisión disminuye al 75% al sustituir estas variables por otras surrogadas, que pueden obtenerse en la práctica clínica de manera no invasiva, como el strain o la presión arterial, con una subida del error relativo hasta el 15% con una probablidad del 95%.
Irrespectively of its origin, diastolic disfunction is present in virtually all structural myocardial diseases. For this reason, it is of high interest to clinicians to understand the changes that occur in the mechanical properties of the heart, with a special interest in the left ventricle, as it is more prone to failure given its high working pressure. Direct models exist that can relate sarcomere properties to chamber properties and their function. However, since sarcomere properties are virtually impossible to measure in vivo or in silico, no model exists to solve the corresponding inverse problem. The present thesis assesses several convolutional neural networks that target sarcomere mechanical properties using different pairs of synthetic hemodynamical variables, which have been obtained with the CircAdapt model, in a transfer learning approach. Results show that, when no white noise is considered, accuracy levels above 90% are observed regardless of the hemodynamical variables used in the training process. When white noise is introduced, LV pressure and volume maintain the same level of accuracy. However, this accuracy drops to 75% when the former variables are substituted by other surrogated ones, which can be measured with non-invasive procedures, such as strain or arterial pressure, with an increase in relative error of the parameters up to 15% with a 95% probability.
Irrespectively of its origin, diastolic disfunction is present in virtually all structural myocardial diseases. For this reason, it is of high interest to clinicians to understand the changes that occur in the mechanical properties of the heart, with a special interest in the left ventricle, as it is more prone to failure given its high working pressure. Direct models exist that can relate sarcomere properties to chamber properties and their function. However, since sarcomere properties are virtually impossible to measure in vivo or in silico, no model exists to solve the corresponding inverse problem. The present thesis assesses several convolutional neural networks that target sarcomere mechanical properties using different pairs of synthetic hemodynamical variables, which have been obtained with the CircAdapt model, in a transfer learning approach. Results show that, when no white noise is considered, accuracy levels above 90% are observed regardless of the hemodynamical variables used in the training process. When white noise is introduced, LV pressure and volume maintain the same level of accuracy. However, this accuracy drops to 75% when the former variables are substituted by other surrogated ones, which can be measured with non-invasive procedures, such as strain or arterial pressure, with an increase in relative error of the parameters up to 15% with a 95% probability.
Descripción
Categorías UNESCO
Palabras clave
Citación
Centro
Facultades y escuelas::Facultad de Ciencias
Departamento
Física Matemática y de Fluídos