Aprendizaje máquina aplicado a la segmentación de imágenes ecográficas de la arteria carótida para la medida del grosor íntima-media

Abstract

[SPA] Las enfermedades cardiovasculares son la principal causa de mortalidad, morbilidad y discapacidad a nivel mundial. Gran parte de estas patologías derivan de la aterosclerosis, una enfermedad que afecta a las arterias de mediano y gran calibre provocando su endurecimiento y pérdida de elasticidad. La aterosclerosis se caracteriza por el engrosamiento de la capa más interna de las paredes arteriales debido al depósito de materia grasa, colesterol y otras sustancias. Por tanto, produce un estrechamiento del lumen arterial dificultando el flujo sanguíneo normal. A largo plazo, puede llevar a una oclusión total del vaso afectado, impidiendo la llegada de oxígeno a la zona irrigada y provocando accidentes cardiovasculares severos. Así, es crucial el diagnóstico precoz de la aterosclerosis con fines preventivos. En este sentido, el grosor íntima-media o IMT (Intima-Media Thickness) de la arteria carótida común se considera un marcador precoz y fiable de la aterosclerosis y, por tanto, del riesgo cardiovascular. Las paredes de los vasos sanguíneos están formadas por tres capas, de la más interna a la más externa: íntima, media y adventicia. El IMT se define como la distancia entre las interfaces lumen-íntima y media-adventicia y es evaluado mediante imágenes ecográficas que muestran un corte longitudinal de la arteria carótida común. Esta modalidad de imagen es no-invasiva para el paciente y relativamente económica, aunque suele ser bastante ruidosa y muy dependiente del operador. Además, el IMT se suele evaluar de forma manual, marcando pares de puntos sobre la imagen. Estos aspectos dan un carácter subjetivo a la medida del IMT y afectan a su reproducibilidad. La motivación de esta Tesis Doctoral es la mejora del proceso de evaluación del IMT sobre ecografías de la arteria carótida común. El objetivo fundamental consiste en explorar y proponer diferentes soluciones basadas en técnicas de Aprendizaje Máquina adecuadas para la segmentación de estas imágenes. De esta forma, se pretende detectar las interfaces lumen-íntima y media-adventicia a nivel de la pared posterior del vaso para medir el IMT sin necesidad de la interacción con el usuario. Este hecho implica que las estrategias propuestas resulten adecuadas tanto para el diagnóstico en la práctica clínica diaria como para facilitar el desarrollo de estudios sobre un gran número de imágenes. En particular, el proceso de evaluación del IMT se lleva a cabo en tres etapas completamente automáticas. En la primera etapa se realiza un pre-procesado de las ecografías para detectar la región de interés, es decir, la pared posterior de la arteria carótida común. Seguidamente, se procede a la identificación de las interfaces que definen el IMT. Por último, una etapa de post-procesado depura los resultados y define los contornos finales sobre los que realizar la medida del IMT. Para la detección automática de la región de interés (ROI) se han estudiado dos propuestas diferentes: una basada en Morfología Matemática y otra basada en Aprendizaje Máquina. Sobre la ROI detectada, la segmentación de las interfaces lumen-íntima y media-adventicia se plantea como un problema de Reconocimiento de Patrones, a resolver mediante técnicas de Aprendizaje Máquina. Así, se han estudiado cuatro configuraciones diferentes, utilizando distintos algoritmos de entrenamiento, arquitecturas, representaciones de los datos de entrada y definiciones del espacio de salida. Por tanto, la segmentación se reduce a realizar una clasificación de los píxeles de la ecografía. El post-procesado ha sido adaptado a cada una de las estrategias de segmentación propuestas para detectar y eliminar los posibles errores de clasificación de forma automática. Una parte importante del estudio realizado se dedica a la validación de las técnicas de segmentación desarrolladas. Para ello, se ha utilizado un conjunto de 79 ecografías adquiridas con el mismo equipo de ultrasonidos, pero utilizando diferentes sondas y con diferentes resoluciones espaciales. Además, se ha realizado la segmentación manual de todas las imágenes por parte de dos expertos diferentes. Considerando como ground-truth el promedio de cuatro segmentaciones manuales, dos de cada experto, se han evaluado los errores de segmentación de las estrategias automáticas planteadas. El proceso de validación se completa con la comparación de las medidas automáticas y manuales del IMT. Para la evaluación de los resultados, se han empleado diagramas de cajas, análisis de regresión lineal, diagramas de Bland-Altman y diferentes parámetros estadísticos. Los procedimientos desarrollados han demostrado ser robustos frente al ruido y artefactos que puedan presentar las ecografías. También se adaptan a la variabilidad anatómica e instrumental de las imágenes, lográndose una segmentación correcta con independencia de la apariencia que muestre la arteria en la imagen. Los errores medios obtenidos son similares, o incluso inferiores, a los errores propios de otros métodos automáticos y semiautomáticos encontrados en la literatura. Además, como consecuencia de utilizar máquinas de aprendizaje, el proceso de segmentación destaca por su eficiencia computacional. [ENG] Cardiovascular diseases are the leading cause of mortality, morbidity and disability worldwide. Large proportion of these diseases results from atherosclerosis, an illness that affects arterial blood vessels causing the hardening and loss of elasticity of the walls of arteries. Atherosclerosis is characterized by the thickening of the innermost layer of the arterial walls due to the accumulation of fatty material, cholesterol and other substances. Therefore, it produces a narrowing of the arterial lumen which hinders the normal blood flow. In the long term, it can lead to an entire occlusion of the affected vessel, preventing the flow of oxygen to the irrigated area and causing severe cardiovascular accidents. Thus, an early diagnosis of atherosclerosis is crucial for preventive purposes. In this sense, the intima-media thickness (IMT) of the common carotid artery is an early and reliable indicator of atherosclerosis and, therefore, of the cardiovascular risk. The walls of blood vessels consist of three layers, from the innermost to the outermost: intima, media and adventitia. The IMT is defined as the distance between the lumen-intima and media-adventitia interfaces and it is assessed by means of ultrasound images showing longitudinal cuts of the common carotid artery. This imaging modality is noninvasive and relatively low-cost, although it tends to be quite noisy and highly operator dependent. Usually, IMT is manually measured by the specialist, who marks pairs of points on the image. These aspects give a subjective character to the IMT measurement and affect its reproducibility. The motivation of this Ph.D. Thesis is the improvement of the evaluation process of IMT in ultrasound images of the common carotid artery. The main objective is the exploration and the proposition of different solutions based on Machine Learning for segmenting these images. In this way, it is intended to detect the lumen-intima and media-adventitia interfaces in the posterior wall of the vessel to measure the IMT without user interaction. This means that the proposed strategies are suitable both for the diagnosis in daily clinical practice and to facilitate the development of studies with a large number of images. In particular, the evaluation process of IMT is carried out in three fully automatic stages. The first stage is a pre-processing of the ultrasound image in which the region of interest (ROI), i.e. the far-wall of the common carotid artery, is detected. Then, it proceeds to the identification of the interfaces defining the IMT. Finally, a post-processing stage debugs the results and defines the final contours on which IMT is evaluated. Two different proposals have been studied for the ROI detection: one based on Mathematical Morphology and the other based on Machine Learning. Once the ROI is detected, the segmentation of the lumen-intima interface and the media-adventitia interface is posed as a Pattern Recognition problem and it is solved by Machine Learning techniques. Thus, four different configurations have been developed by using distinct architectures, training algorithms, representations of input information and output space definitions. Therefore, segmentation is reduced to perform a classification of the pixels belonging to the ROI. The post-processing stage has been adapted to each one of the proposed segmentation strategies to detect and eliminate possible misclassifications in an automatic way. An important part of the present study is dedicated to the validation of the developed techniques. For this purpose, 79 images acquired with the same ultrasound equipment, but using different probes and different spatial resolutions, have been used. Two experts have performed the manual segmentation of all the images. Considering as ground-truth the average of four manual segmentations, two from each expert, the segmentation errors of the four different strategies have been evaluated. The validation process is completed with the comparison between automatic and manual IMT measurements. For an exhaustive characterization of the results, box plots, linear regression analysis, Bland-Altman plots and different statistical parameters have been used. Developed procedures have proven to be robust against noise and artifacts that may appear in the ultrasounds. They also adapt themselves to the anatomical and instrumental variability of the images, achieving a correct segmentation regardless of the appearance of the artery in the ultrasound. The obtained mean errors are similar, or even lower, than errors in other automatic and semi-automatic methods. Moreover, as a result of using learning machines, the segmentation process stands out for its computational efficiency.

[ENG] Cardiovascular diseases are the leading cause of mortality, morbidity and disability worldwide. Large proportion of these diseases results from atherosclerosis, an illness that affects arterial blood vessels causing the hardening and loss of elasticity of the walls of arteries. Atherosclerosis is characterized by the thickening of the innermost layer of the arterial walls due to the accumulation of fatty material, cholesterol and other substances. Therefore, it produces a narrowing of the arterial lumen which hinders the normal blood flow. In the long term, it can lead to an entire occlusion of the affected vessel, preventing the flow of oxygen to the irrigated area and causing severe cardiovascular accidents. Thus, an early diagnosis of atherosclerosis is crucial for preventive purposes. In this sense, the intima-media thickness (IMT) of the common carotid artery is an early and reliable indicator of atherosclerosis and, therefore, of the cardiovascular risk. The walls of blood vessels consist of three layers, from the innermost to the outermost: intima, media and adventitia. The IMT is defined as the distance between the lumen-intima and media-adventitia interfaces and it is assessed by means of ultrasound images showing longitudinal cuts of the common carotid artery. This imaging modality is noninvasive and relatively low-cost, although it tends to be quite noisy and highly operator dependent. Usually, IMT is manually measured by the specialist, who marks pairs of points on the image. These aspects give a subjective character to the IMT measurement and affect its reproducibility. The motivation of this Ph.D. Thesis is the improvement of the evaluation process of IMT in ultrasound images of the common carotid artery. The main objective is the exploration and the proposition of different solutions based on Machine Learning for segmenting these images. In this way, it is intended to detect the lumen-intima and media-adventitia interfaces in the posterior wall of the vessel to measure the IMT without user interaction. This means that the proposed strategies are suitable both for the diagnosis in daily clinical practice and to facilitate the development of studies with a large number of images. In particular, the evaluation process of IMT is carried out in three fully automatic stages. The first stage is a pre-processing of the ultrasound image in which the region of interest (ROI), i.e. the far-wall of the common carotid artery, is detected. Then, it proceeds to the identification of the interfaces defining the IMT. Finally, a post-processing stage debugs the results and defines the final contours on which IMT is evaluated. Two different proposals have been studied for the ROI detection: one based on Mathematical Morphology and the other based on Machine Learning. Once the ROI is detected, the segmentation of the lumen-intima interface and the media-adventitia interface is posed as a Pattern Recognition problem and it is solved by Machine Learning techniques. Thus, four different configurations have been developed by using distinct architectures, training algorithms, representations of input information and output space definitions. Therefore, segmentation is reduced to perform a classification of the pixels belonging to the ROI. The post-processing stage has been adapted to each one of the proposed segmentation strategies to detect and eliminate possible misclassifications in an automatic way. An important part of the present study is dedicated to the validation of the developed techniques. For this purpose, 79 images acquired with the same ultrasound equipment, but using different probes and different spatial resolutions, have been used. Two experts have performed the manual segmentation of all the images. Considering as ground-truth the average of four manual segmentations, two from each expert, the segmentation errors of the four different strategies have been evaluated. The validation process is completed with the comparison between automatic and manual IMT measurements. For an exhaustive characterization of the results, box plots, linear regression analysis, Bland-Altman plots and different statistical parameters have been used. Developed procedures have proven to be robust against noise and artifacts that may appear in the ultrasounds. They also adapt themselves to the anatomical and instrumental variability of the images, achieving a correct segmentation regardless of the appearance of the artery in the ultrasound. The obtained mean errors are similar, or even lower, than errors in other automatic and semi-automatic methods. Moreover, as a result of using learning machines, the segmentation process stands out for its computational efficiency.