Model-Based Analysis of Cerebrovascular Diseases Combining 3D and 4D MRA Datasets

Abstract

Detailed knowledge about the vascular anatomy and blood flow at a macro- and microvascular level is often required for diagnosis, therapy decision, and intervention in case of cerebrovascular diseases. This knowledge can be obtained from high resolution 3D and time-resolved magnetic resonance imaging datasets. However, advanced computer-assisted image analysis methods are needed for an improved and faster diagnosis of patients with a cerebrovascular disease, due to the massive amount of acquired data. From a methodical perspective, an automatic cerebrovascular segmentation, hemodynamic analysis, and combined visualization of vascular structures together with the corresponding hemodynamic situation is required. The multi-step segmentation framework presented in this work was developed to enable a sufficient delineation of all vessels, including small and malformed vessels, from high resolution 3D angiographies. The purpose of the first step of this segmentation framework is to extract the brain tissue using a graph-based approach to prevent leakage of the segmentation into non-brain tissue in the subsequent processing steps. After this, a non-linear combination of intensity and shape information using fuzzy logic is performed, in which the shape information is obtained from the vesselness filter. The resulting parameter image is then used for an automatic extraction of the cerebrovascular system employing a level-set segmentation approach with anisotropic energy weights. Finally, gaps in the resulting cerebrovascular segmentation are reduced applying a shortest path algorithm. The automatically segmented vessels can, for example, be used for a surface-based visualization, which enables an intuitive screening of the anatomy. However, this visualization includes no information about the blood flow. Therefore, time-resolved magnetic resonance angiographies have to be analyzed. The direct estimation of hemodynamic parameters from the corresponding concentration time curves is often imprecise. This drawback can be overcome using the developed reference-based linear curve fitting approach. Here, the main idea is to extract a patient-individual hemodynamic reference curve directly from the given time-resolved dataset. The extracted reference curve is fitted to each concentration time curve and then used for parameter estimation. Finally, a novel 4D blood flow visualization technique was developed, which enables a direct combined analysis and visualization of the cerebrovascular system and the corresponding blood flow situation. Therefore, the 3D and 4D MRA image sequences of a patient are registered such that the extracted hemodynamic information can be displayed dynamically over time on the vessel surface model. An evaluation of the cerebrovascular segmentation framework using manual segmentations as ground truth showed that this approach is capable of segmenting small as well as malformed vessels with high precision within the range of inter-observer differences, and significantly better than typical state-of-art methods. Furthermore, a comparison to commonly applied hemodynamic models using Monte-Carlo simulations and clinical datasets revealed that the proposed reference-based linear curve fitting model leads to reduced errors regarding the time-to-peak parameter over a wide range of temporal resolutions. Finally, an evaluation of the 4D blood flow visualization using datasets of patients with arteriovenous malformations showed that this visualization leads to a feasible representation of the cerebrovascular anatomy and blood flow, which allows an intuitive and fast diagnosis. The developed methods have been already successfully applied for several clinical studies in addition to the 4D blood flow visualization. Within this context, the hemodynamic situation in the presence of anatomical rupture risk factors of arteriovenous malformations was analyzed and a computer-aided segmentation and angiographic characterization of this vascular pathology was developed. Further examples for clinical applications described in the second part of this thesis comprise the improved measurement of aneurysms, quantitative analysis of aneurysm coil treatment stability, as well as tissue-at-risk quantification and identification of patients within 4.5 hours of symptom onset in case of an acute ischemic stroke. In conclusion, the developed methods may help to improve the understanding, diagnosis, and treatment of cerebrovascular diseases.