Automation and Evaluation of Image-Derived Workflows on 3D and 4D Cardiac CT

Cardiovascular diseases are a major public health concern and currently the leading cause of mortality and morbidity worldwide. Medical imaging supports several cardiovascular disease applications, and recent technological advancements have increased the quality and quantity of medical images available. In turn, the additional information provided has been supporting diagnosis procedures, treatments and patient monitoring. However, to access the contained information, images need to be manipulated to a certain extent. The manipulation normally starts with segmentation or landmarking procedures that serve as an input for further processes such as patient-specific biomechanical simulation or 3D-printing. These procedures are still tedious, time-consuming, error-prone and user-dependent. The recent explosive growth in image quantity and quality has moreover increased the resources required to complete these manipulation processes. There is, therefore, a strong need for reliable, robust and automated tools to address the limitations of the current processes. The work in this thesis addresses these limitations by automating and evaluating current image-derived workflows. This work focuses on three specific topics: segmentation, landmarking and biomechanical simulation. Given the relevance of cardiovascular diseases, and the widespread use of Computed Tomography (CT), the scope of this thesis will be limited to 3D and 4D cardiovascular CT. Image segmentation is a process used to label the structure of interest of the cardiovascular anatomy and often represents the first step of an image processing workflow. Many technologies have attempted to simplify or to automate the process; however, manual input is still the norm. In this work, a novel automatic method for the segmentation of cardiac CT is designed, developed and validated. The method is based on an efficient combination of atlas-registration and graph-cut technology. The validation is performed with 95 cardiac CTs, and the results show comparable accuracy and unmatched computational speed in comparison with other automated tools. The annotation of crucial anatomical structures, also known as landmarking, is a process that is often performed to morphologically assess anatomical structures. In 4D-CT, the process is especially relevant to monitor changes in the mitral annulus dimension, in order to support improvements in the treatment of patients suffering from mitral regurgitation. The annotation is however still mostly performed manually, making the process especially tedious. This work shows how deformable image registration can be used to automate landmark annotation by propagating landmarks from one phase of the image stack to all other phases. The accuracy of the method was evaluated by measuring the distance between the propagated and manually annotated landmarks on 8 4D-CTs. The results support the use of this technology as a way to automate the current landmark annotation process. Finally, the prediction of peak stress of the aortic wall, obtained through biomechanical simulations, represents a promising tool to support aortic aneurysm risk stratification. The simulations require the use of robust and biofidelic models based on patient-specific information. Therefore, mechanical properties are a crucial input, but patient-specific information is not available in clinical cases. In this work, we evaluate the possibility to estimate the material properties \textit{in vivo} starting from 4D-CT, and as such, improve the prediction of peak wall stress. The accuracy of the method is rigorously evaluated by performing a virtual experiment. The results support the use of patient-specific material properties to perform biomechanical simulations and show that the method leads to an accurate estimation of the material properties and of the consequent peak wall stress only if the material model used provides a true representation of the material properties of the aorta. Moreover, the evaluation platform developed in the process can be easily expanded to support future developments and improvements of the method. Altogether, this thesis presents significant engineering contributions to three image-processing workflows stemming from image acquisition: segmentation, landmarking, and biomechanical simulations. These improvements enable the extraction of information from cardiac CT more easily and efficiently. Hence, they have the potential to provide clinicians and researchers with additional resources and information by removing the need to address imaging-data processing tasks. Moreover, the additional information could provide a more comprehensive understanding about the cardiac anatomy and about the behavior of soft tissues.

Jaar van publicatie:2020

Toegankelijkheid:Open