|Title:||Automatic venous vessel segmentation in high field, multi-echo susceptibility weighted imaging||Language:||English||Authors:||Rechberger, Albert||Qualification level:||Diploma||Keywords:||MRT; Multi-echo Susceptibility Weighted Imaging; Venous Vessel Segmentation||Advisor:||Badurek, Gerald||Issue Date:||2016||Number of Pages:||90||Qualification level:||Diploma||Abstract:||
Susceptibility-Weighted Imaging (SWI) is a Magnetic Resonance Imaging (MRI) technique that exploits both the magnitude and phase of the complex MRI signal to increase contrast for tissues of different susceptibilities. Deoxygenated blood in in venous vessels is more paramagnetic than the surrounding tissues, therefore veins can be depicted in SWI without the need for external contrast agents. Identifying and segmenting the venous vessels in whole-brain SWI scans facilitates the creation of three-dimensional models of the cerebral venous vasculature. However, manual segmentation of veins in whole brain SWI datasets is unfeasible due to the amount of manual labor required. To date automatic segmentation approaches of veins from ultra-high field SWI datasets have predominately been performed using only the magnitude images because of the non-local and orientation dependent properties of the phase. However, in recent years, dedicated algorithms have been established which aim to turn the complex phase information into maps of the local susceptibility, a process that is known as Quantitative Susceptibility Mapping (QSM). In this project, a new approach to automatic venous vessel segmentation was developed that uses information from magnitude images, phase images and the derived QSM images of a multi-echo T_ 2 -weighted gradient echo scan. A Random Forest (RF) classifier was used to segment veins based on a combination of appearance and shape features that are computed separately from magnitude images, phase images and QSM images. This supervised machine-learning approach also allowed us to investigate the importance of each feature for the segmentation task. This not only gives insight to the importance of magnitude, phase and QSM images for venous vessel segmentation, but because the features were computed from multiple echoes, the feature importance findings can also be used to suggest echo time settings for future data acquisition. The segmentation approach was tested on datasets of five different healthy subjects, two of which were partially annotated to serve as ground truth for training the RF and for quantitatively evaluating the segmentation performance. In all of the performance metrics used within our experiments, the RF approach yielded higher scores than either of those features used individually. Specifically, the RF approach outperformed the common vesselness filtering approach in all similarity measures that were computed against the manual annotations. Visual assessment of 3D renderings of the surface veins confirmed that the segmentations obtained by the RF approach did look very similar to renderings of the manual annotations. The results of the feature importance measurements indicate that most of the information that is needed for surface vein segmentation is already contained in the first echo, which potentially enables quicker data acquisition. Overall the developed RF segmentation approach enables the generation high-quality, patient-specific 3D models of the cerebral venous vasculature, which have the potential to aid neurosurgeons in presurgical planning by helping them to localize brain regions that need to be spared in order to minimize the risk of post-operative neurological deficits.
|Library ID:||AC13373508||Organisation:||E141 - Atominstitut||Publication Type:||Thesis
|Appears in Collections:||Thesis|
Show full item record
Files in this item:
|Automatic venous vessel segmentation in high field multi-echo susceptibility weighted imaging.pdf||28.62 MB||Adobe PDF|
checked on Feb 18, 2021
checked on Feb 18, 2021
Items in reposiTUm are protected by copyright, with all rights reserved, unless otherwise indicated.