AI Driven End-to-End 4D flow MRI Pipeline for Hemodynamic Assessment in BAV Disease - SUMMARY / ABSTRACT Bicuspid aortic valve (BAV) is the most common congenital heart defect (1-2% in the U.S.) which can lead to severe complications including aortic dilatation or dissection. However, the current risk-stratification paradigm is based on primitive aortic diameter thresholds and has poor predictive value for outcomes. Studies have shown that 4D flow MRI can measure in-vivo complex 3D blood flow as risk factors for aortic complications in BAV. At Northwestern University (NU), 4D flow MRI has been used for hemodynamic assessment of the aorta since 2011 and we have established a large database with over 4,000 aorta 4D flow MRI scans in >2,000 unique patients and >300 healthy controls. Using this data we have identified promising hemodynamic indices for risk stratification in BAV. For example, elevated WSS was found to be a novel biomarker for vessel wall architectural remodeling and identifies patients at risk of progressive aortic dilation. However, broad clinical adoption of 4D flow MRI is hindered by long scan times (8-15 minutes) and cumbersome analysis, limiting reproducibility and translation. To overcome these challenges, we developed deep learning (DL) models to highly accelerate 4D flow MRI acquisition and automate analysis workflows. Our goal is to build on these findings to establish an end-to-end 4D flow MRI DL pipeline for highly efficient data acquisition and hemodynamic analysis with full on-scanner deployment. The large NU database of manually processed 4D flow MRI data (n=4,000+) provides unique ground truth data for training and testing of DL workflows. To account for variability across sites and MRI systems, DL developments will be augmented by aorta 4D flow data by our subaward sites, chosen to allow for validation across the main MRI hardware vendors (NU: Siemens, UW: GE, CU: Phillips). The final DL networks and processing tasks will be embedded in containerized applications to allow for multi-vendor on-scanner deployment as well as dissemination and prospective multi-site evaluation. We aim to 1) rigorously optimize and validate deep learning accelerated 4D flow MRI for the accurate measurement of aortic hemodynamics; 2) utilize automated hemodynamic analysis to conduct a retrospective large cohort study with known 10-year BAV outcomes, and 3) conduct a prospective multi-center and multi- vendor evaluation study of the end-to-end 4D flow MRI deep learning pipeline. This study will introduce innovative DL methods for 4D flow MRI, assess the role of hemodynamics in BAV 10-year outcomes, and validate the multi-site feasibility and performance of the end-to-end 4D flow pipeline.
