Motion Resistance Artificial Subtraction Angiography with Deep Learning

Catheter Digital Subtraction Angiography (DSA) is an imaging technique that was developed in the 1980s to allow physicians to visualize blood vessels. Today, this technology is utilized for minimally-invasive interventions that treat numerous devastating pathologies, including stroke and myocardial infarction, diseases that disproportionally impact underserved minority patient populations. DSA is performed by inserting a catheter into an artery, injecting iodinated contrast, and recording a series of X-Ray images as the contrast traverses the patient’s blood vessels. However, superimposed X-Ray densities from bones and soft tissues obscure the imaging details of the blood vessels. In ideal conditions, DSA will provide an image of the vessels alone, unobscured by superimposed bone and soft tissue. Indeed, during angiography of cooperative awake patients, who are instructed to hold their breath to reduce motion, DSA can produce excellent images. However, DSA images are markedly degraded by all voluntary, respiratory, or cardiac motion that occurs during the exam. In situations where patients are unable to remain still, which may be due to difficulty breathing or the distress of an acute stroke, the poor quality of motion-degraded DSA imaging increases the risk of complex procedures such as stroke clot removal and cardiac stenting. We have developed a deep learning algorithm that can perform the task of DSA even in the setting of substantial motion. We utilize a 3D U-Net neural network architecture, which is optimized to use the spatial and temporal information in the images to identify the blood vessels and separate them from the other X-ray densities such as bone and soft tissue. In this grant application, we will this innovative algorithm to the bedside. The algorithm will be implemented on the Picture Archiving and Communications System (PACS), where physicians can evaluate the results of our technology side-by-side with DSA on real patient data. As the second major goal of this proposed work, we will optimize the algorithm to perform low-latency, real-time inference, which will be needed for implementation in live procedures. At the end of the funding period, we will deliver a validated clinical implementation of our Deep Learning Angiography algorithm, as well as an optimized version of our algorithm for real-time use during X-ray guided interventions, which will be integrated into angiography machines in future work.