2575. AI-Derived Systolic 3D Aortic Hemodynamics from Aortic Geometry
Authors * Denotes Presenting Author
  1. Haben Berhane *; Northwestern University
  2. Anthony Maroun; Northwestern University
  3. Ulas Bagci; Northwestern University
  4. Michael Markl; Northwestern University
  5. Bradley Allen; Northwestern University
Quantification of aortic flow is vital for diagnosis and patient management. While 4D Flow MRI provides a comprehensive assessment of aortic hemodynamics, it requires long acquisition times, cumbersome preprocessing, and is not widely available. Alternatively, computational fluid dynamics (CFD) is capable of simulating 3D blood flow dynamics from aortic geometry. However, it is hampered by requiring user-defined boundary conditions, long simulation times, and patient-specific in-flow and pressure conditions. Deep learning has recently shown success in image-to-image translation. As such, we sought to develop a CycleGAN framework to derive systolic 3D blood flow velocities in the aorta from 3D aortic segmentation data as the only input.

Materials and Methods:
This study used a total of n = 1765 patients (1242 bicuspid aortic valve [BAV] patients, median age: 42 years; 523 trileaflet aortic valve [TAV] patients, median age:45 years) with aortic 4D flow MRI acquired on either 1.5T or 3T MRI systems (Siemens). 4D flow MRI sequence parameters were: spatial res=1.2-5.0mm3, venc=150-500cm/s. A 3D aortic segmentation was generated from the 4D flow and used to extract the 3D distribution of systolic velocities inside the aorta. For AI training (n=994 BAV, n=419 TAV), 3D aorta segmentations served as input to estimate 3D systolic velocities, and the ground truth data was 4D flow derived aortic 3D velocities. AI testing was performed on 248 BAV and 104 TAV patient data that were not included in training. The CycleGAN was composed of two generators and two discriminators[2]. For the generators, we used a 3D hybrid Densenet/Unet[3]. For the discriminators, we used 5 convolution layers of a kernel size=[4x4]. Two separate CycleGANs were trained: one for BAV datasets and another for TAV. Maximum intensity plots (MIPs) were generated for both AI-derived and 4D flow 3D velocities. Region of interests were used to quantify peak velocities (top 5%) in the ascending (AAo), arch, and descending aorta (DAo). Statistical comparisons were performed using a paired t-test or Wilcoxon signed rank test based on normality.

Across all testing datasets, the AI successfully estimated systolic aortic 3D velocities with strong agreement compared to the 4D flow MRI. Voxel-wise comparison of AI vs. 4D flow 3D systolic velocities showed a strong correlation between the methods (ICC = 0.93, low bias: 0.01m/s) and good limits of agreement (±0.3m/s). For regional peak velocity analysis, correlation demonstrated excellent agreement across all three regions (AAo, arch, DAo, ICC = 0.98-0.99 [0.96-0.98 - 0.98-0.99]. Bland-Altman analysis showed small bias for all regions (AAo: 0.03m/s, arch: 0.01m/s, DAo: 0.01m/s for both datasets) and good to moderate limits of agreement: AAo: ±0.22-0.27m/s, arch: ±0.13-0.14m/s, DAo: ±0.15m/s. No significant differences were found in the peak velocities across all datasets (AI: 1.27±0.63, 4D: 1.28±0.64).

AI-based peak systolic velocity estimation showed strong agreement to the 4D flow MRI. Future direction is to develop velocity estimations from CT-derived 3D aortic geometry.