Nicolas Blondel

Semester Work
Supervisors: Neerav Karani, Prof. Ender Konukoglu

Segmentation of the aorta in 4D flow MRIs without ground truth labels

4D flow magnetic resonance (MR) imaging is an emerging technique that allows acquisition of time resolved 3D MR images, coupled with quantitative blood flow information. This allows for measurement and visualization of temporal evolution of complex blood flow patterns within the acquired 3D volume. Assessment of such flow information can potentially aid in prediction of adverse events such as aneurysms, as well as for long-term prognosis of interventions such as valve replacement. Towards such automatic assessment, the focus of this thesis is the automation of segmentation of the aorta from 4D flow MRIs. Recent successes of learning-based approaches in medical image segmentation often rely on the availability of relatively large amounts of annotated data. Noting the impracticality of obtaining expert annotations on 4D images, we develop an interactive segmentation tool based on the random walker algorithm. The tool leverages intensity as well as flow similarity between neighbouring voxels (in space and time) and provides a 4D segmentation in less than a couple of minutes. Additionally, it allows for repeated user feedback to improve its performance. Using this tool, we assemble a non-expert-annotated dataset and train a neural network to obtain 4D segmentations on new images in a few seconds, while not requiring any user interaction. While quantitative evaluation of the predicted segmentations is not possible in the absence of ground truth annotations, qualitative inspection shows that they are adequately accurate and will provide an excellent basis for further analysis - for instance, for anomalous blood flow identification in 4D flow MRI data.