3D reconstruction of coronary arteries using deep networks from synthetic X-ray angiogram data

• Published in: Communications Faculty of Sciences University of Ankara Series A2-A3 Physical Sciences and Engineering Vol. 64; no. 1; pp. 1 - 20
• Main Authors: ATLI, İbrahim, GEDİK, Osman Serdar
• Format: Journal Article
• Language: English
• Published: 30.06.2022
• ISSN: 2618-6462; 2618-6462
• DOI: 10.33769/aupse.1020175

Cardiovascular disease (CVD) is one of the most common health problems that are responsible for one-third of all deaths around the globe. Although X-Ray angiography has deficiencies such as two-dimensional (2D) representation of three dimensional (3D) structures, vessel overlapping, noisy background, the existence of other tissues/organs in images, etc., it is used as the gold standard technique for the diagnosis and in some cases treatment of CVDs. To overcome the deficiencies, great efforts have been drawn on retrieval of actual 3D representation of coronary arterial tree from 2D X-ray angiograms. However, the proposed algorithms are based on analytical methods and enforce some constraints. With the evolution of deep neural networks, 3D reconstruction from images can be achieved effectively. In this study, we propose a new data structure for the representation of objects in a tubular shape for 3D reconstruction of arteries using deep learning. Moreover, we propose a method to generate synthetic coronaries from data of real subjects. Then, we validate tubular shape representation using 3 typical deep learning architectures with synthetic X-ray data we produced. The input to deep learning architectures is multi-view segmented X-Ray images and the output is the structured tubular representation. We compare results qualitatively in terms of visual appearance and quantitatively in terms of Chamfer Distance and Mean Squared Error. The results demonstrate that tubular representation has promising performance in 3D reconstruction of coronaries. We observe that convolutional neural network (CNN) based architectures yield better 3D reconstruction performance with 9.9e-3 on Chamfer Distance. On the other hand, LSTM-based network fails to learn the coronary tree structure and we conclude that LSTMs are not appropriate for auto-regression problems as depicted in this study.