Deep Learning To Automate Late Mechanical Activation Detection From Cardiac Magnetic Resonance Images

点击翻译按钮获取中文摘要

Project summary: This proposal aims to develop advanced machine learning and artificial intelligence (ML/AI) techniques to rapidly and accurately identify sites with late mechanical activation (LMA) and compute circumferential uniformity estimate with singular value decomposition (CURE-SVD) from standard cine cardiac magnetic resonance (CMR) images. Our long-term goal is to develop networks that can determine LMA sites / CURE-SVD automatically from cine images acquired at any CMR facility worldwide, thereby addressing a critical need in the effective guidance of device- based therapies, such as Cardiac resynchronization therapy (CRT), for potentially millions of heart failure patients. To accomplish this goal, we will make use of a rich and unique dataset we have assembled at our institution based on over 200 patients undergoing CRT with a median follow-up of five years. The data set includes demographics and comorbid diseases from EHR review, pre- CRT/post-CRT imaging with CMR cine/DENSE/LGE (late gadolinium enhancement), echocardiography, and multidimensional response parameters based on overall survival, serum B-type natriuretic peptide testing, quality of life questionnaires, and exercise testing for peak VO2. The central hypothesis of this proposal is that these ML/AI methods will effectively identify the characteristics of scar-free LMA sites from cine imaging, achieving excellent agreement compared with the original DENSE-based assessments, and predict post-CRT outcomes. Our specific aims are (i) identifying LMA sites and computing CURE-SVD by developing joint neural networks with inputs from cine SSFP/GRE images, (ii) with the addition of scar from LGE in the network, we will develop a novel multi-task learning to consider scar information in the determination of LMA sites free of scar, and (iii) comparing the performance of our proposed methods with ground truth DENSE and results obtained from commercial feature tracking software to predict CRT outcomes in the dataset with 200+ CRT patients with complete CRT response data and long-term follow-up for survival and arrhythmia outcomes.