A novel approach in 3D reconstruction of coronary artery tree from 2D X-ray angiograms

Cardiovascular diseases are common health problems in both developing and developed countries and responsible for one-third of all deaths around the globe. Therefore, accurate diagnosis of vascular diseases plays a critical role in decreasing the rate of deaths and improving the quality of life of the whole world. Although there are many imaging modalities used to diagnose coronary diseases such as DSA, MRA, CTA, etc., X-Ray angiography is used as a gold standard technique among them in clinics. However; it is limited by inherent two-dimensional (2D) representation of three-dimensional (3D) structures and the diagnosis relies heavily on the experience of cardiologists. In this manner, 3D imaging technologies provide an objective, operator-independent tool for an accurate assessment, especially in visualization and quantification of blood vessels. In addition, this technology has the potential to guide clinical decisions. Therefore, the development of an automated and accurate vessel-tree reconstruction from angiograms is highly desirable. In literature, several methods exist for 3D reconstruction of coronary artery trees, but they propose analytical methods that require camera calibration or feature extraction to match correspondences in images for the utilization of epipolar geometry properties. Some of these methods require manual assistance. However; recent developments in technology and tremendous effort drawn in learning-based 3D reconstruction methods enable objects to be reconstructed without any camera calibration or feature extraction steps. In this thesis, we develop an end-to-end fully automated pipeline using deep learning architectures for 3D reconstruction of coronary artery tree from X-Ray angiograms. The pipeline contains mainly 2 steps: (1) blood vessel segmentation and (2) 3D reconstruction of vessels segmented in the first part. We propose a novel fully convolutional deep learning architecture, called Sine-Net, for blood vessel segmentation and multiple deep learning architectures for 3D reconstruction. The input to 3D reconstruction networks is the segmented image extracted in the first part, and the output is a well-defined 3D representation of connected cylinders. This structured definition for 3D tubular shapes is novel to the literature. As deep learning architectures need plenty of training data with ground-truth, we populate 3D coronary arterial trees synthetically from real data of 9 subjects. After generating relatively enough data for training, we propose multi-view CNN, LSTM and GRU based architectures to predict 3D model of given multiple segmented vessel images. We have validated our method and the data structures defined for 3D tubular shapes in the 3D reconstruction of coronary arteries.