AbstractCardiovascular disease (CVD) is one of the leading causes of death and disability in the UK. There is therefore much interest in the early staging of CVD to help improve patient outcomes through long-term treatments. Whole-body magnetic resonance angiography (WBMRA) offers a quantitative assessment of whole body arterial atheromatous disease burden, making it a useful diagnostic tool for CVD. However the large datasets produced are very labour-intensive to examine, and there is an unmet need for automated software tools to help clinicians with their diagnoses. Towards this goal, this thesis proposes an automatic framework for the processing and analysis of WBMRA data for grading stenoses. This includes registration of pre- and post-contrast volumes, automatic vessel segmentation and tracking, and the measurement and grading of tracked vessels.
We present the first quantitative comparison of vessel segmentation techniques for WBMRA data, comparing five different algorithms using 3 ground truth vessel maps annotated manually following a clear protocol. We find that a U-Net convolutional neural network algorithm outperforms previous, well established algorithms despite the limited amount of training data.
To enable the development and validation of stenosis grading algorithms, we also gathered ground truth stenosis annotations for 18 patients with three trained clinicians. A thorough analysis of the inter- and intra-rater variability revealed a higher disagreement between annotators than expected for manually detecting and grading stenoses, which is largely unexamined in the literature. The development of clear protocols for the collection of this ground truth data, in close collaboration with clinical partners, enables us to present good-practice guidelines for ground truth collection from WBMRA data for algorithm development.
Finally, we present three stenosis detection algorithms tested against synthetic vessels and well-characterised ground truth stenosis annotations.
|Date of Award
|Manuel Trucco (Supervisor), Graeme Houston (Supervisor) & Ian Poole (Supervisor)