Deep learning-based image registration in dynamic myocardial perfusion CT imaging

Registration of dynamic CT image sequences is a crucial preprocessing step for clinical evaluation of multiple physiological determinants in the heart such as global and regional myocardial perfusion. In this work, we present a deformable deep learning-based image registration method for quantitative myocardial perfusion CT examinations, which in contrast to previous approaches, takes into account some unique challenges such as low image quality with less accurate anatomical landmarks, dynamic changes of contrast agent concentration in the heart chambers and tissue, and misalignment caused by cardiac stress, respiration, and patient motion. The introduced method uses a recursive cascade network with a ventricle segmentation module, and a novel loss function that accounts for local contrast changes over time. It was trained and validated on a dataset of n = 118 patients with known or suspected coronary artery disease and/or aortic valve insufficiency. Our results demonstrate that the proposed method is capable of registering dynamic cardiac perfusion sequences by reducing local tissue displacements of the left ventricle (LV), whereas contrast changes do not affect the registration and image quality, in particular the absolute CT (HU) values of the entire CT sequence. In addition, the deep learning-based approach presented reveals a short processing time of a few seconds compared to conventional image registration methods, demonstrating its application potential for quantitative CT myocardial perfusion measurements in daily clinical routine.