Intracranial vascular abnormalities and related diseases affect millions of people each year. Accurate and reliable extraction and visualization of intracranial vessels are therefore essential in the clinical setting. Vessel extraction of segmentation may enhance the diagnosis of the disease, its status assessment and longitudinal monitoring, and surgery planning and preparation. With the advancement of automatic methods based on machine learning models, the process of vessel segmentation is less time-consuming and, at the same time, more accurate and reliable.
In this thesis, we explored the field of vascular extraction with deep learning methods, such as the U-Net convolutional neural network and a publicly accessible IXI database, which contains 570 3D magnetic resonance angiography (MRA) scans.
For evaluation purposes, we have manually segmented the intracranial vessels in these scans. Additionally, in about 80 scans the circle of the Willis region was annotated.
In the experiments, we performed an objective and quantitative evaluation and cross-comparison of state-of-the-art vessel segmentation methods. We proposed a novel approach based on localized U-net model learning anchored to the circle of Willis since there is the highest incidence of intracranial aneurysms. According to the overlap between the reference and obtained automatically segmentations quantified using the Dice-Sorensen coefficient (DSC), the proposed approach to intracranial vascular segmentation shows significantly higher values compared to the current state-of-the-art approaches.
The average DSC values were computed using cross validation and obtained a DSC of 0.9478, which exceeds the values of the state-of-the-art result by more than 1 %. For research purposes and further validation of vessel segmentation methods, we also prepared a database called IXI-Angio, which contains the 570 manually created reference segmentations of the intracranial vessels in MRA images of the IXI dataset.
|
