Image analysis and deep learning to support endovascular repair of abdominal aortic aneurysms
An abdominal aortic aneurysm (AAA) is a focal dilation of the aorta that may lead to its rupture. The most common treatment for AAAs is endovascular aneurysm repair (EVAR). EVAR implies lifelong post-operative surveillance using Computed Tomography Angiography (CTA), due to the potential appearance of complications. This thesis sets the basis for intelligent CTA image analysis to support post-operative follow-up of AAAs, providing clinicians with valuable information to prognose the behavior of the aneurysm.
First, novel pre-operative and post-operative AAA segmentation approaches are developed, based on Convolutional Neural Networks (CNN). Initially, 2D AAA detection and segmentation CNNs are proposed. Then, segmentation is extended to 3D to increase segmentation accuracy. Precise AAA segmentation is the basis for a good AAA follow-up. It allows to measure aneurysm volume, which is thought to be a better indicator for aneurysm rupture than the current AAA diameter measurements. Furthermore, it enables more complex analyses of AAA morphology and deformations.
Subsequently, a methodology for post-operative CTA time-series registration and aneurysm biomechanical strain analysis is also proposed. From these strains, quantitative image-based descriptors are extracted and correlated with the long-term patient prognosis. The extracted descriptors are the basis for possible future imaging biomarkers to be used in clinical practice to assess patient prognosis and to enable informed decision making after EVAR.
Finally, the technological developments in the thesis are applied to solve complex segmentation problems in other clinical domains, such as pectoral muscle segmentation from mammograms and pulmonary artery segmentation from CT scans.
Validation of the 3D AAA segmentation approach proposed in this thesis is being carried out with the aim of integrating it in a commercial product.