Automatic Cerebral Perfusion Imaging using Deep Learning for Digital Subtraction Angiography

Abstract

Visualization of blood flow in brain vessels is crucial for neurovascular disease patients, including the evaluation of ischemic stroke treatments. X-ray Digital Subtraction Angiography (DSA) is the standard imaging modality used for this purpose. So far, visual inspection is the primary way to assess DSA series. Due to the high temporal resolution of DSA, it holds great potential in facilitating quantitative assessment of cerebral hemodynamics. Various parametric perfusion images have been generated from DSA based on temporal blood flow characteristics. This technique is commonly referred to as perfusion DSA. Such parameters include cerebral blood volume (CBV), cerebral blood flow (CBF), time to maximum (Tmax), and mean transit time (MTT) and can be generated using deconvolution techniques. To obtain these deconvolution-based images, an arterial input function (AIF) is extracted from the internal carotid artery (ICA) region of interest. However, current perfusion DSA methods require manual annotation of the ICA. In this work, a supervised deep-learning model for semantic ICA segmentation was trained. Subsequently, an automated application was developed to generate perfusion DSA images. The results of quantitative statistics indicate that there is no significant difference between using the automated application that utilizes the AI segmentation model and the manual annotation method. These findings suggest that this application framework may offer important benefits in clinical practice and future research.