Next-Generation Cardiovascular MRI powered by Artificial Intelligence

点击翻译按钮获取中文摘要

Project Summary/Abstract: Despite the accuracy and versatility of cardiovascular MRI, its footprint is only 1% among cardiac imaging tests (SPECT, echocardiography, CT, MRI) in the US. While there are several factors such as referral patterns favoring SPECT and echocardiography among cardiologists that account for low utilization, the two addressable obstacles that preclude widespread adoption are lengthy scan time (imaging facility operational cost) and reading (physician cost). These obstacles must be addressed for community hospitals with limited resources to adopt cardiovascular MRI into clinical routine practice. While compressed sensing (CS), since its introduction into the MRI world in 2007, has led to highly-accelerated cardiovascular MRI acquisitions, the subsequent image reconstruction remains too slow (> 5 min for 2D time series, > 1 hour for 3D time series) for clinical translation (unmet need 1). Downstream, image analysis for cardiovascular MRI is notoriously labor intensive (e.g. 30- to 60-min) and limited (“circles” at two cardiac phases for cine MRI, whereas perfusion and late gadolinium-enhanced (LGE) images are evaluated visually), for what is essentially a basic computer vision task (unmet need 2). In direct response, we will address these two unmet needs and unlock the enormous potential of CMR using deep learning (DL). DL applications have exploded since advancements in optimization and GPU hardware. While several recent studies have applied neural networks such as convolutional neural networks (CNNs), U-Nets, and Generative Adversarial Nets (GANs) for reconstruction and segmentation, no study has implemented an inline end-to-end pipeline that receives raw k-space from the MRI scanner and delivers both reconstructed images and fully processed images automatically with high speed (< 1 min). The objectives of this study are: a) developing a network for image reconstruction with maximal acceleration (aim 1), (b) developing a network for image processing tasks (aim 2), and c) developing an integrated, end-to-end network that does both (aim 3). By developing an architecture that can simultaneously learn maximal acceleration, fine tune end-to-end performance, and perform reconstruction/inference using feed-forward networks, we anticipate a disruptive technology that will lead to a paradigm shift in cardiovascular MRI and increase its footprint in community hospitals. This 2-year study is doable because of the requisite database of raw k-space (not derived from DICOM) data (N = 617) and annotated cardiac MR images (N=3,021) from over 3,000 patients existing at our institution. Success of this proposal will deliver a disruptive technology that has potential to cause a paradigm shift in cardiovascular MRI and enable widespread adoption of cardiovascular MRI into clinical routine practice.