Segmentation of vascular structures based on adaptive tracking and new seed detection

Abstract

In this dissertation, methods to construct vascular structures in computed tomography angiography (CTA) are proposed. The construction is useful in numerous applications, including vascular surgery, interventional radiology (IR), and medical education. The construction commences with tracking of the central axis of a vessel segment. In this dissertation two adaptive tracking methods are proposed to extract a cerebral arterial segment in CTA. Segmentation of cerebral arteries in CTA is very challenging mainly due to bone contact and vein contamination. Both the two proposed tracking methods iteratively accumulate elliptical cross-sections from the given initial seed. The first method is very intuitive. The boundary of each cross-section is modelled as an ellipse with border points detected by gradient thresholds. The thresholds are established adaptively to bone contact and vein contamination. This adaptiveness improves robustness of the tracking compared to a conventional method. The other tracking method is augmented by rigorous mathematical framework of a particle filter (PF). Its measurement model is defined as the border points on the plane normal to the axis as the observation. The border points are detected using the radial component of the gradient and the adaptive thresholds in the same way as the intuitive tracking. The outliers in the measurement vector are discarded in the computation of the distance between each particle and the actual observation. The robustness of the PF tracking method is confirmed by intensive experiments on both clinical and synthesized vessel data. Since vascular structures are composed a number of vessel segments, tracking of a vessel segment need to be supplemented by any methods to extract many vessel segments. Specifically to cerebral arteries in CTA, the head CT volume is partitioned into two sub-volumes according to the amount of contact between bone and arteries. In the lower sub-volume, as the arteries are often adjacent to bone or veins, they are extracted by the above-mentioned tracking methods. In the upper sub-volume where bone and vessels are not adjacent to each other, vessels can be segmented well by intensity thresholds and connected component analysis (CCA). Separate application of different algorithms to lower and upper sub-volumes leads to discontinuity in the partitioning slice and inter-partition tracking algorithm is also proposed to resolve this discontinuity problem. The other method to construct vascular structure is based on detection of new seeds and can be generalized to other vascular structures, i.e. pulmonary vasculature, cardio-vasculature or hepatic vasculature. New seed is detected by searching for tubular object near the vessel segment already extracted. Objects are classified into tubular objects or non-tubular objects according to the eigenvalues of the covariance matrix. Construction results of the two methods are compared to Registration-Subtraction results of pre-contrast and post-contrast scanned volumes and are evaluated to show comparable performance by a clinical expert. Finally, as an application of the proposed methods, coronary arteries are extracted, which is not much different from extraction of cerebral arteries. It can be estimated as easier, since the main two problems of bone contact and vein contamination need not be considered. However, limited contrast resolution may cause leakage of the tracking. As multi-detector row CT (MDCT) technology develops, non-invasive image of coronary arteries becomes realizable and their extraction becomes more important.