Automatic Segmentation of Coronary Arteries from Computed Tomography Angiography Data Using Optimal Thresholding

Abstract

This thesis presents an automated framework for delineation of coronary arteries from computed tomography angiography data cloud. Delineation of coronary artery is an important component in assisting and automating different radiologist task. Generally, manual analysis of huge amount of data generated by latest imaging modalities such as CTA is time consuming and timely interpretation of such data is necessary for diagnosing the life threatening diseases. Despite intensive research, efficient segmentation of coronary arteries from medical imaging data remains an active research area. There is no single segmentation algorithm that works well due to various factors that may include imaging modality, intensity inhomogeneity etc. Therefore, the proposed framework introduced in this dissertation presents an automated method to segment the coronary arterial tree by exploiting the idea of optimal thresholding. Generally, medical images suffer from intensity inhomogeneity problem which makes it difficult to use a single threshold to successfully delineate the region of interest. Especially, when the task is to segment the coronary arteries from CTA data, the use of a single threshold would not be sufficient due to a large number of slices having varying intensities. Therefore, in this dissertation a novel coronary segmentation method has been proposed that successfully segments a pair of coronary arteries by determining a globally optimal threshold. The method first generates the vessel probability map known as vesselness measure by analyzing the Hessian values to cope with the intensity inhomogeneity problem. Hessian based vesselness map gives the likelihood of each voxel to belong to vascular structures. For handling vessels of different sizes, the vesselness map is computed at different scales and each voxel is assigned the strongest likelihood value. The vessel probability map is further refined in the later stage by determining the local and global optimal thresholds. The local optimal threshold is computed for each slice of a given CTA volume, whereas the global optimal threshold is determined by computing the mean of all the local optimal thresholds. Finally, a fraction of global threshold is used to segment the complete coronary arterial structure accurately. In various segmentation algorithms, the delineation procedure commence with the user provided initial point. However, in such cases the outputs of the algorithms mainly relies on the level of user interaction. To get rid of this dependency on user interaction and to make the segmentation process automatic, the proposed framework in this dissertation exploits the anatomical relationship between aorta and coronary arteries. It is known that aorta is a major vessel within the human body hence can be used as an indicator to guide the delineation procedure of other connected structures such as coronary arteries. Generally, coronary arteries initiate from the ascending aorta as the two largest objects connecting the aorta. Therefore, the proposed method detects the aorta using Hough Transform technique by considering the circular anatomy of the aorta. After detection of the aorta, the coronary arteries are identified by extracting the two largest components connected with the aorta. The performance of the proposed framework is validated on real clinical CTA and publicly available datasets in terms of efficiency and accuracy. Experimental results reveal the capability of the proposed method in extracting both of the coronary arteries completely without leakages and gaps. Furthermore, the method is compared with the two previous approaches as well as with the manually delineated ground truth data obtained by radiologists. A high degree of accuracy of the proposed framework has been observed when compared with the previous methods and the ground truth data. On average, the proposed method achieves 95% accuracy, whereas Lankton’s and Khedmati’s method attain 65% and 50% accuracy as compared to the ground truth data.