Skip to main content
2024 BMES Annual Meeting Main banner
Icon Legend
This session is not in your schedule.
This session is in your schedule. Click again to remove it.

Biomedical Imaging and Instrumentation

Biomedical Imaging and Instrumentation - Poster Session E

Poster R7 - Geometric Validation of Deep Learning Segmentation-Based Numerical Mesh Generation of Aortic Valve from CT Images

Saturday, October 26, 2024
10:00 AM - 11:00 AM EST
Location: Exhibit Hall E, F & G

Introduction:  Aortic stenosis (AS) elevates mortality risk for 300,000 Americans annually due to aging or congenital valve issues. Transcatheter aortic valve replacement (TAVR) has overtaken surgical aortic valve replacement as the standard for AS treatment. Though TAVR has been shown to be as effective as the surgical counterpart while remaining minimally invasive, certain risks with TAVR are still prevalent: paravalvular leakage, conduction abnormalities, increased risk of stroke, and myocardial tears. However, these risks are mitigated through robust pre-procedural planning which includes precise anatomic measurement and size selection via CT imaging. Pre-procedural planning can be improved through the usage of computational simulations, such as FEA, CFD, and FSI, to determine proper device type and size. To utilize simulations, 3D numerical meshes of patient-specific cardiac anatomy is required; but this requires time and expertise to create such meshes, thus limiting the clinical viability of the method. Deep learning (DL) algorithms improve mesh generation speed and effectiveness, but lack comparison and testing for complete TAVR planning workflows. This study demonstrates the geometric validity of the meshes generated by DeepCarve, one such DL algorithm, and sets a protocol to compare the models of other cardiac digital twin generating models.

Materials and

Methods:  3 patients who have undergone TAVR with pre-procedural CT imaging at Stony Brook University were randomly selected and the patient-specific DL-generated meshes were created using DeepCarve and manually segmented models were created using ScanIP. The meshes were aligned to original CT coordinates from metadata with manual refinements, and they were then cut to a standardized height. Cropped meshes were re-meshed to 0.1 to 0.2 mm facet lengths. Centerlines for both mesh types were extracted using 3DSlicer and their Procrustes distance as well as point-to-point distances were calculated. These were compared against a null hypothesis of perfect congruence for significance. Along the centerline, slices were cut at standardized heights with respect to the interpoint angular axis and the areas of each slice between mesh types were compared. Lastly, point cloud metrics were used to investigate the bodily similarity of the aorta as well as just the isolated sinotubular junction through Chamfer and Hausdorff distance and principal component analysis.

Results, Conclusions, and Discussions: Across all patients, there were noted inconsistencies in size and shape, notably within the annulus and sinotubular junction regions between the DL-generated meshes and manually segmented meshes. The DL-generated meshes exhibited constricted, non-physiological contours in this region, where geometric deviations in this region are critical in terms of affecting fidelity of FEA and FSI simulations. When comparing the centerlines, analysis reveals the closest Procrustes distance (PD) alignment in patient 1 (PD = 0.0021) and the most pronounced divergence is in patient 3 (PD = 0.1009), which reflects the degree of visual discrepancy. This would notably impact the deployment of self-expandable TAVRs, where the impact would be seen less for balloon expandable variants. Patient 2's meshes aligned most closely in cross-sectional area comparisons, while Patients 1 and 3's DL meshes overestimated much of the cross-sectional areas significantly (p < 0.05). Patient 3's DL mesh exhibited the most overestimation with an average overestimation of 8.7%. These area discrepancies may influence the validity of fluid dynamics simulations within the DL vessel meshes. Lastly, patient 2's meshes demonstrated the highest point cloud integrity (Chamfer distance = 0.133 mm), while Patient 3's models showed the least (Chamfer distance = 0.253 mm). A similar trend followed for the Hausdorff distances. Increased Chamfer distances were noted when just analyzing an isolated sinotubular junction region. Furthermore, PCA of the meshes indicated that highest variability in the x and z directions of the aortas (who were standing upright towards the positive z-direction), suggesting anisotropic geometric fidelity across the meshes. Overall, geometric dissimilarities will likely lead to diverging simulation results, thus limiting viability for clinical use. In the near future, FEA and fluid simulations will be run with self-expandable and balloon-expandable TAVR models in order to capture the true viability of DL generated meshes. There is work being done to update the current DeepCarve model to create a less constricted sinotubular junction geometry, which will then also be compared with manually segmented meshes and older DeepCarve DL meshes with the same protocol.

Acknowledgements (Optional):