Vessel-diameter quantification and embolus detection in CTA images

Abstract

Pulmonary embolism (PE) is the sudden obstruction of an artery in the lungs, usually due to a blood clot. There are more than 50 cases of PE per 100,000 persons every year in the USA. Of these cases, 11% die in the first hour and in total, the untreated mortality rate of PE is estimated to be 30%. Thus, PE is a common disorder with a high morbidity and mortality for which an early and precise diagnosis is highly desirable. Contrast-enhanced multi-slice x-ray computed tomography (CT) has become the preferred initial imaging test (and often the only test) to diagnose PE, because it is a simple, minimally invasive, fast and high-resolution imaging technique that allows the direct depiction of a clot inside the blood vessels. The CT image can also be used to identify other potentially life-threatening causes in a patient with chest pain. In contrast-enhanced CT (i.e., CT angiography, CTA) images, the blood vessels appear to be very bright because the contrast material is dissolved in blood. The embolus does not absorb contrast material, and therefore, it remains darker. PE can be recognized in CTA images as a dark area in the pulmonary arteries. However, manual detection of the dark spots that correspond to PE in CT images is often described by radiologists as difficult and time consuming. Therefore, computer-aided diagnosis (CAD) is desirable. In this thesis, we propose a new CAD system for automatic detection of PE in CTA images. The evaluation shows that the performance of our system is at the level of state of the art literature. The data was selected to demonstrate a variety of thrombus load, considerable breathing artifacts, sub-optimal contrast and parenchymal diseases, and none of the emboli were excluded for evaluation. This is important because the main problem of PE detection is the separation between true PE and look-alikes, which is much harder when the patient is not healthy. The CAD system that we propose consists of several steps. In the first step, pulmonary vessels are segmented and PE candidates are detected inside the vessel segmentation. The candidate detection step focusses on the inclusion of PE – even when vessels are completely occluded – and the exclusion of false detections, such as lymphoid tissue and parenchymal diseases. Subsequently, features are computed on each of the candidates to enable classification of the candidates. The feature-computation step does not only focus on the intensity, shape and size of an embolus, but also on relative locations and the regular shape of the pulmonary vascular tree. In the last step, classification is used to separate candidates that represent real emboli from the other candidates. The system is optimized with feature selection and classifier selection. Several classifiers have been tested and the results show that the performance is optimized by using a bagged tree classifier with the features distance-to-parenchyma and stringness. The system was trained on 38 CT data sets. Evaluation on 19 other data sets showed that the system generalizes well. The sensitivity of our system on the evaluation data is 63% at 4.9 false positives per data set, which allowed the radiologist to improve the number of detected PE with 22%. Another part of this thesis is about the accurate quantification of the vessel diameter in CT images. Quantification of the local vessel diameter is essential for the correct diagnosis of vascular diseases. For example, the relative decrease in diameter of a stenosis is an important factor in determining the treatment therapy. However, inherent to image acquisition is a blurring effect, which causes a bias in the diameter estimation of most methods. In this thesis, we focus on fast and accurate (unbiased) vessel-diameter quantification. For the localization of the vessel wall, Gaussian derivatives are often used as differential operators. We show how these Gaussian derivatives should be computed on multi-dimensional data with anisotropic voxels and anisotropic blurring. The voxels and blurring are usually anisotropic in the 3D CT images, which means that the voxel size and the amount of blur is not equal in all three directions. Furthermore, we show that the computational cost of interpolation and differentiation on Gaussian blurred images can be reduced by using B-spline interpolation and approximation, without losing accuracy. We introduce a derivative-based edge detector with unbiased localization on curved surfaces in spite of the blur in CT images. Finally, we propose a modification of the full-width at half-maximum (FWHM) criterion to create an unbiased method for vessel-diameter quantification in CT images. This criterion is not only cheaper but also more robust to noise than the commonly used derivative-based edge detectors.