1345235 - AI Derived Systolic 3D Aortic Hemodynamics from Aortic Geometry

Quantification of aortic flow is vital for patient management in a number of heart valve diseases. While 4D Flow MRI provides a comprehensive assessment of aortic hemodynamics, it requires long acquisition times, cumbersome pre-processing, and is not widely available. In contrast, Computational Fluid Dynamics (CFD) is capable of simulating 3D blood flow from aortic geometry but is hampered by requiring user-defined boundary conditions, long simulation times, and patient-specific in-flow and pressure conditions [1]. Recently, artificial intelligence (AI) has shown success in image-to-image translation [2]. As such, we developed a new CycleGAN framework for the prediction of systolic 3D blood flow velocity vector fields with 3D aortic geometry data as the only input.

Methods:

This study used a total of 1765 aortic 4D flow MRI data (1242 bicuspid aortic valve [BAV] patients, median age:42 years; 523 trileaflet aortic valve [TAV] patients, median age: 45 years), acquired on either 1.5T or 3T MRI systems (Siemens). All 4D flow MRI data were acquired with 3D coverage of the thoracic aortic and the following parameters: spatial res=1.2-5.0mm³, venc=150-500cm/s, temp res=32.8-44.8ms. For all subjects, a 3D aortic static anatomic segmentation (used as CycleGAN input) was generated from the 4D flow and used to extract the systolic 3D velocity vector field inside the aorta (ground-truth data). Two separate CycleGANs were trained using a 80/20% split: one for BAV (training: N=994; testing: N=248) and another for TAV (training: N=419; testing: N=104). The CycleGAN was composed of two generators (3D hybrid Densenet/Unet) and discriminators (series of 5 convolution layers) [2-4] (Figure 1). Region of interest analysis was used to quantify peak velocities (top 5%) in the ascending (AAo), arch, and descending aorta (DAo). Comparisons were performed using maximum intensity projections (MIPs) and 3D vector field plots as well as voxel-by-voxel and regional peak velocity Bland-Altman analysis.

Results:

AI computation time was 0.15±0.11s, while total training time was 3600 min. Figure 2 illustrates AI performance for one of the best (A) and worst (B) cases. Figure 2A shows excellent qualitative agreement between AI and 4D flow velocity patterns, such as vortex flow in the AAo. Figure 2B shows one of the worst examples in which the AI-derived velocities resulted in an incorrect jet-flow pattern in AAo traveling along the opposite side of the aorta. Nonetheless, voxel-by-voxel Bland Altman comparisons showed strong AI vs 4D flow agreement for both cases. Bland-Altman analysis of regional peak velocities are shown in Figure 3 and demonstrate small bias for all regions (AAo: 0.04-0.05 m/s, arch: -0.04m/s, DAo: -0.03-0.04m/s for both datasets) and strong limits of agreement: AAo: ±0.22-0.41m/s, arch: ±0.14-0.24m/s, DAo: ±0.12-0.15m/s (Table 1).

Conclusion:

AI-based systolic velocity estimation showed strong agreement to the 4D flow MRI. Future direction is to expand this work for CT 3D aortic geometry.