Automatic Curvature Analysis of the Left Atrium from Cardiac Magnetic Resonance Imaging

Pla Alemany, Sofia. Automatic Curvature Analysis of the Left Atrium from Cardiac Magnetic Resonance Imaging. 2022, Master Thesis, University of Basel, Faculty of Medicine.

[img] PDF
Restricted to Repository staff only


Official URL: https://edoc.unibas.ch/88255/

Downloads: Statistics Overview


Atrial Fibrillation (AF) is one of the most common heart arrhythmias, affecting about 1 to 2% of the general population. Given that the AF trigger is often initiated by pulmonary veins (PVs), pulmonary vein isolation (PVI) is the cornerstone of interventional treatments for AF today. Although it has proven efficiency and safety, AF recurrence (AFR) frequently occurs posterior to a PVI. Therefore, a careful assessment of patients who would benefit from this procedure has to be considered individually. In the clinic, the selection of patients is mainly based on clinical factors (i.e. type of AF, age, or previous AF). New image segmentation and computer vision techniques have allowed the study of the correlation between shape features of the heart and their relationship with AFR, allowing an optimized patient selection and to the anticipation PVI interventions’ success. Nevertheless, there is still an incomplete understanding of all the factors that serve as potential biomarkers for AFR prediction.
AF’s activation is caused by the uncontrolled automaticity of certain cells in the heart. Among other factors, this could potentially result from tissue inhomogeneities and sharp curvatures of the tissue that leads to a pronounced stretch of cells. Therefore, LA curvature features might be considered a surrogate for increased myocardial stretch and hence AFR. Consequently, curvature quantification could potentially be valuable to investigate the clinical significance of the mechanical stretch in the LA and PVs of patients who underwent catheter ablation.
This study aims to develop an automatic, consistent framework to acquire global and local geometrical features of the left atrium (LA) and the first approximately 2 cm of the pulmonary veins (PVs) of patients that underwent PVI. We train an autoencoder (AE) network to segment the object of interest from preprocedural cardiac Magnetic Resonance Imaging (cMRI) scans. We developed a pipeline that aims to reconstruct its 3D surface and locate certain regions on the LA and PVs that could be of high interest to study the correlation between AFR and curvature values in an automated way. These regions of interest (ROIs) include the left and right carinas between the PVs, the ridge between the LA and the LA appendage, and the left and right PV ostium. The location of these ROIs for any new patient is based on an image registration approach and will allow the extraction of local curvature features. Additionally, the registration step is followed by an active contours-based fine-tuning step. The volume, surface area and shape index are also acquired from the entire 3D reconstruction. The local and global features will enable the visualization of statistical results to attain an initial notion of the potential correlation between these features and AFR.
The results of our study show that convolutional AEs are able to segment the LA ad PVs from cMRI scans. Moreover, it proved that a workflow based on image registration and active contours achieves promising results on automatic ROI location on the surface of an anatomical object. Nevertheless, further developments are required to improve the results obtained for subjects with shape anomalies and to automatically detect error locations without visual inspection. Finally, due to the lack of a large data set available for a deep statistical analysis, no strong correlations were found between the extracted features and AFR.
Advisors:Cattin, Philippe and Knecht, Sven
Committee Members:Schnider, Eva
Faculties and Departments:03 Faculty of Medicine > Departement Biomedical Engineering > Imaging and Computational Modelling > Center for medical Image Analysis & Navigation (Cattin)
UniBasel Contributors:Knecht, Sven
Item Type:Thesis
Thesis Subtype:Master Thesis
Thesis status:Complete
edoc DOI:
Last Modified:27 Apr 2022 04:30
Deposited On:26 Apr 2022 09:32

Repository Staff Only: item control page