Multimodal Image Analysis for Carotid Artery Plaque Characterization

Atherosclerosis of the carotid artery is a main cause of ischemic cerebrovascular events. There is evidence that the composition of the vessel wall is more strongly related to plaque vulnerability and subsequent events than luminal stenosis, which is currently used for risk stratification in clinical practice. Noninvasive imaging can characterize the composition of the vessel wall. In order to incorporate measures of plaque composition into clinical practice, accurate and robust image segmentation methods are required.

This thesis describes the development and validation of image analysis techniques that aim at the automated characterization of the carotid atherosclerotic vessel wall. The first part of this thesis makes use of a dataset in which ex vivo and in vivo MRI and CT, and annotated histology sections are available and have been spatially aligned. We firstly perform segmentation of plaque components in ex vivo MRI. Voxel classifiers are trained on a ground truth of registered histology and μCT images. We show the importance of different groups of features: intensities, Gaussian filters and wall distances, and use these features in subsequent work on in vivo data. Here we address the problems that arise in training and evaluation of segmentation methods when misregistration between histology and in vivo data occurs. Still, we show that accurate segmentation of the lipid-rich necrotic core, calcification and fibrous tissue is possible when MRI and CTA are combined, and linear discriminant analysis is performed after rejecting outliers from the training set. Finally, in this first part of the thesis we develop a method for automatic segmentation of different plaque components from histology sections, to make the use of histology for training and evaluation more feasible and less time-consuming. Subsequently we perform plaque component segmentation from in vivo MRI only, and address the fact that MRI datasets acquired in difference centers using different hardware varies considerably in appearance. Firstly, we show that segmentation of lipid, intraplaque hemorrhage, calcification, and fibrous tissue can be performed with similar accuracy as the variation between observers on MRI data from two different centers. Secondly, we show that the accuracy decreases when a method developed on data from one center is used to apply to data from the other center. We propose two methods by which we improve this transferability of segmentation methods: non-linear feature scaling, and transfer learning in which we add only a few annotated slices from the ‘new’ center to the training data. Lastly, we perform a study on texture analysis of carotid artery plaques in 3D ultrasound images. From a large set of texture parameters we obtain the strongest parameters to form a ‘risk indicator’. In a longitudinal study with 3D ultrasound imaging at two time points, we show that change in texture is a stronger predictor of vascular events than previously used parameters for risk stratification, and that using texture in addition to those parameters improves risk stratification in patients with carotid artery disease.