Generative Adversarial Networks for Coronary CT Angiography Acquisition Protocol Correction with Explicit Attenuation Constraints

Rudolf Leonardus Mirjam Van Herten, Louis van Harten, Nils Planken, Ivana Isgum

Keywords: Contrast media timing, coronary attenuation, coronary CT angiography, deep learning, generative adversarial networks

Abstract: The image quality of coronary CT angiography (CCTA) is important for the correct diagnosis of patients with suspected coronary artery disease, which is heavily influenced by image acquisition. Timing of the contrast media injection specifically influences the level of arterial enhancement, and it is aimed to allow optimal assessment of the coronary artery morphology. However, a consensus on an optimal acquisition protocol that can account for the large variety in patient cohorts has not been reached, commonly resulting in suboptimal arterial enhancement. In this work, we propose a generative adversarial network for the retrospective correction of contrast media attenuation in CCTA, thus reducing the dependency on an optimal timing protocol at acquisition. We develop and evaluate the method in a set of 1,179 CCTA scans with varying levels of contrast enhancement. We evaluate the consistency of intensity values in the coronary arteries and evaluate performance of coronary centerline extraction as a commonly performed analysis task. Results show that correction of contrast media attenuation values in CCTA scans is feasible, and that it improves the performance of automatic centerline extraction. The method may allow improved analysis of coronary arteries in CCTA scans with suboptimal contrast enhancement.