Publication Details

Warning: All material on this website, including papers, text, figures, and graphics is covered by Copyright © unless otherwise stated. You may browse them at your convenience (in the same spirit as you may read a journal or a proceeding article in a public library). Retrieving, copying, or distributing these files, however, may violate the copyright protection law. We recommend that the user abides international law in accessing this directory.
infoAbstract pdfPaper/Preprint bibBibTeX doiDOI plusSupplementary Material slidesSlides posterPoster
[All] → [Schneider2014b]
Teaser
 
Teaser

Matthias Schneider. Reconstruction of Complete Cerebrovascular Networks from High-Resolution Tomographic Data. PhD thesis, ETH Zurich, Zurich, Switzerland, Dec 2014.

ETH Medal Award 2016 for Outstanding Doctoral Thesis

Online Version44.58 MBDownload [pdf]
Print Version43.95 MBDownload [pdf]
BibTeX5.99 KBDownload [bib]
Addendum170.97 KBDownload [pdf]
The cerebrovascular system continuously delivers oxygen and nutrients to the brain. Inadequate oxygen supply (cerebral hypoxia) leads to fatal consequences ranging from impairment of the brain function to brain death. As a matter of fact, neurovascular diseases such as stroke are the leading cause of serious long-term adult disability. Likewise, many neurodegenerative disorders such as Alzheimer’s are associated with alterations of the vascular system that can contribute to neuronal loss due to oxygen deprivation. An in-depth understanding of the microvascular structure is thus important to better comprehend pathophysiological cerebral processes and the interpretation of noninvasive functional imaging relying on signals that originate from the vasculature. Similarly, precise knowledge of the cerebrovascular topology is required for numerical modeling of cerebral blood flow dynamics and its regulation in order to improve our understanding of how local disruptions of cerebral blood flow impact the local and global perfusion of brain tissue. Ex-vivo imaging of the entire cerebrovascular system has become feasible with recent developments in high-resolution imaging such as synchrotron radiation X-ray tomographic microscopy (SRXTM). Even if detailed image data of reasonably large samples can be acquired down to the capillary level, there are yet no means available today to comprehensively analyze the imaged complex vascular network. This thesis aims to develop practical and effective methods for the reconstruction of complete cerebrovascular networks from partially incomplete high-resolution angiographic image data. The correction of topological defects due to missing image information, particularly at the capillary level, represents a major challenge in this context. As part of this project, we propose a processing pipeline for vascular reconstruction that involves optimized correction of topological discontinuities. For this purpose, we have developed different algorithms combining conventional methods for vessel segmentation with generative approaches that rely on morphological and functional properties of the cerebral vasculature. We have focused on three main aspects: (1) First of all, we propose a novel machine learning framework for joint 3-D vessel segmentation and centerline extraction based on oblique random forests and multivariate Hough regression using the same set of features. We use steerable filters for the efficient computation of local image features at different scales and orientations in order to account for the multi-scale nature of vascular structures and to compensate for the preferred local vessel orientation. Additionally, we introduce and comprehensively test a novel oblique split model with an elastic net penalty term that imposes sparsity on the optimal split weights and generates shallow decision trees, which results in more efficient prediction. In a series of extensive validation experiments, we systematically evaluate the most important structural components of our approach and demonstrate the advantage of the learning step that allows for highly accurate vessel segmentation and centerline extraction even for automatically generated noisy training data. (2) The second problem investigated in this thesis is the generation of synthetic vascular structures. We present a stochastic modeling approach to generate optimized arterial trees based on physiological principles, while at the same time certain morphological properties are enforced during construction. A simplified angiogenesis model serves as the driving force of vascular morphogenesis and degeneration incorporating case-specific information about the metabolic demand of the tissue in the considered domain. The construction process further includes morphometrically confirmed bifurcation and branch length statistics to optimize the synthetic vasculature. The proposed method is able to generate artificial - yet physiologically plausible - arterial trees that match the metabolic demand of the embedding tissue and fulfill the prescribed morphological properties at the same time. (3) The final focus of this thesis is the automated correction of vascular connectivity which is key to reconstructing consistent vascular networks from partially incomplete image data. Based on the physiologically motivated simulation framework for the generation of synthetic vascular trees, we pursue a generative approach for data-driven topology correction. This approach allows for reconstruction of complete vascular networks that explain the image data and, at the same time, feature topological and morphological consistency. The proposed method leverages both topological and morphological evidence extracted from the image data to iteratively improve and recover vessel connectivity. Thus, small topological discontinuities with strong support by the configuration of proximate vessel stubs can be identified and overcome, while more complex subnetworks are synthesized to correct larger defects lacking satisfactory image-based evidence.
@PhDThesis{Schneider2014b, title = {Reconstruction of Complete Cerebrovascular Networks from High-Resolution Tomographic Data}, author = {Matthias Schneider}, school = {ETH Zurich}, year = {2014}, address = {Zurich, Switzerland}, month = {Dec}, abstract = {The cerebrovascular system continuously delivers oxygen and nutrients to the brain. Inadequate oxygen supply (cerebral hypoxia) leads to fatal consequences ranging from impairment of the brain function to brain death. As a matter of fact, neurovascular diseases such as stroke are the leading cause of serious long-term adult disability. Likewise, many neurodegenerative disorders such as Alzheimer’s are associated with alterations of the vascular system that can contribute to neuronal loss due to oxygen deprivation. An in-depth understanding of the microvascular structure is thus important to better comprehend pathophysiological cerebral processes and the interpretation of noninvasive functional imaging relying on signals that originate from the vasculature. Similarly, precise knowledge of the cerebrovascular topology is required for numerical modeling of cerebral blood flow dynamics and its regulation in order to improve our understanding of how local disruptions of cerebral blood flow impact the local and global perfusion of brain tissue. Ex-vivo imaging of the entire cerebrovascular system has become feasible with recent developments in high-resolution imaging such as synchrotron radiation X-ray tomographic microscopy (SRXTM). Even if detailed image data of reasonably large samples can be acquired down to the capillary level, there are yet no means available today to comprehensively analyze the imaged complex vascular network. This thesis aims to develop practical and effective methods for the reconstruction of complete cerebrovascular networks from partially incomplete high-resolution angiographic image data. The correction of topological defects due to missing image information, particularly at the capillary level, represents a major challenge in this context. As part of this project, we propose a processing pipeline for vascular reconstruction that involves optimized correction of topological discontinuities. For this purpose, we have developed different algorithms combining conventional methods for vessel segmentation with generative approaches that rely on morphological and functional properties of the cerebral vasculature. We have focused on three main aspects: (1) First of all, we propose a novel machine learning framework for joint 3-D vessel segmentation and centerline extraction based on oblique random forests and multivariate Hough regression using the same set of features. We use steerable filters for the efficient computation of local image features at different scales and orientations in order to account for the multi-scale nature of vascular structures and to compensate for the preferred local vessel orientation. Additionally, we introduce and comprehensively test a novel oblique split model with an elastic net penalty term that imposes sparsity on the optimal split weights and generates shallow decision trees, which results in more efficient prediction. In a series of extensive validation experiments, we systematically evaluate the most important structural components of our approach and demonstrate the advantage of the learning step that allows for highly accurate vessel segmentation and centerline extraction even for automatically generated noisy training data. (2) The second problem investigated in this thesis is the generation of synthetic vascular structures. We present a stochastic modeling approach to generate optimized arterial trees based on physiological principles, while at the same time certain morphological properties are enforced during construction. A simplified angiogenesis model serves as the driving force of vascular morphogenesis and degeneration incorporating case-specific information about the metabolic demand of the tissue in the considered domain. The construction process further includes morphometrically confirmed bifurcation and branch length statistics to optimize the synthetic vasculature. The proposed method is able to generate artificial - yet physiologically plausible - arterial trees that match the metabolic demand of the embedding tissue and fulfill the prescribed morphological properties at the same time. (3) The final focus of this thesis is the automated correction of vascular connectivity which is key to reconstructing consistent vascular networks from partially incomplete image data. Based on the physiologically motivated simulation framework for the generation of synthetic vascular trees, we pursue a generative approach for data-driven topology correction. This approach allows for reconstruction of complete vascular networks that explain the image data and, at the same time, feature topological and morphological consistency. The proposed method leverages both topological and morphological evidence extracted from the image data to iteratively improve and recover vessel connectivity. Thus, small topological discontinuities with strong support by the configuration of proximate vessel stubs can be identified and overcome, while more complex subnetworks are synthesized to correct larger defects lacking satisfactory image-based evidence.}, doi = {10.3929/ethz-a-010340252}, keywords = {vascular reconstruction, vessel segmentation, centerline extraction, vascular connectivity, synthetic vascular network, cerebral vasculature, computational physiology, angiogenesis, oxygen metabolism, oblique random forest, multivariate Hough voting, steerable filters}, url = {http://e-collection.library.ethz.ch/view/eth:47141} }