Artery-vein separation from thoracic CTA scans with application to PE detection and volume visualization
MetadataShow full item record
The automatic reconstruction and separation of arterial and venous vascular trees is an essential step for automatic computer-aided diagnostic systems used for maladies such as pulmonary embolism, coronary artery disease, hepatic cirrhosis, brain arteriovenous malformation and intracranial hemorrhage. Automatic artery-vein (AV) reconstruction and separation still remains a challenging problem in computational imaging and geometry processing due to various reasons: low signal to noise ratio in the images, varying intensity of similar vascular tissue, varying contrasts, imaging artifacts, motion blurring, and unclear boundaries between arteries and veins due to low image resolution. In this dissertation, we present an automatic and robust method for AV separation in 3D thoracic CT-Angiography (CTA) images. In our method, the heart, lungs, muscles, and blood vessels are first classified and segmented using intensitybased expectation maximization and morphological operations. Next, the dista transform is used to find the boundary of blood vessels and the heart. Using a tracing procedure, we convert the 3D CTA image into a graph representation. Next, by detecting the pulmonary trunk, and via a graph traversal, we connect blood vessel branches to the pulmonary trunk from the end of vessels and leading toward the heart, while merging and verifying branches to solve the cases where there are unclear boundaries. If a branch is linked to the pulmonary trunk, then all subsegmental trees of the branch are classified as pulmonary arteries. Otherwise, it is classified as a pulmonary vein when it is connected to a ventricles/auricles of the heart, or it is a disconnected branch when there is no connection to the heart. We validate our automatic AV separation algorithm with topology matches with the bronchial tree for different CTA imaging cases. As an application of our AV separation method, we also provide an automatic pulmonary embolism (PE) detection algorithm. Pulmonary embolism is defined as a thrombus (blood clot) that forms in a part of the venous system, becomes detached, and lodges in a pulmonary artery after passing through the right chambers of the heart. Untreated PE is potentially lethal and should be promptly diagnosed and treated with anticoagulants. However, PE diagnosis from CTA is often difficult for several reasons: high volume of image slices to be interpreted, manual/visual pulmonary artery separation from veins, contrast changes, and many similar features in images. Our PE detection method is fully automatic and identifies the locations of PE in 3D CTA images. We have calibrated and verified our approach with manually diagnosed CTA imaging data sets. Finally, we present a method for volumetric image visualization. We present vii a multi-dimensional transfer function (histogram of values, gradient magnitudes) that contains spatial information and enables visual differentiation of features even when two different features overlap in the 2D histogram. Furthermore, we provide details of an implementation of our transfer function on modern programmable graphics hardware.