Session: 12-26-01: Data-Driven Modeling and Simulation for Computational Biomedicine

Paper Number: 71442

Start Time: Friday, 03:45 PM

71442 - A Computational Pipeline for Generating Dynamic, High-Order, Patient-Specific Meshes for Use in Cardiac Biomechanics Simulations 

Computational modeling of the heart plays a vital role in understanding normal cardiac function, as well as abnormal behavior due to various cardiovascular diseases. Accurate and successful cardiac biomechanics simulations require patient-specific geometric models of the heart. High-order partial differential equation solvers, such as finite element methods, have the potential to perform such simulations, provided these solvers are paired with a high-order mesh that faithfully captures the highly-curved boundaries and various features of the heart.

In this presentation, we will describe the computational pipeline we have developed to generate dynamic, patient-specific, feature-preserving, high-order tetrahedral cardiac meshes from cardiac magnetic resonance (CMR) images.  To extract the cardiac motion, we rely on image registration to find an optimal flow between consecutive 3D frames of a 4D cine CMR dataset.  To this end, we propose a convolutional neural network-based 4D deformable registration technique for consistent motion estimation [1].  Our unsupervised deep learning framework employs a Laplacian-based operator as smoothing loss for deformable registration of 3D cine CMR images.  Prior to registration, the images are corrected for slice misalignment, which is common during CMR image acquisition.  We segment the left and right ventricles from the end-diastole image and generate corresponding surface meshes, then decimate them in Meshlab to reduce the number of triangular faces.  We then convert the low-order surface mesh to a high-order surface mesh using MeshCurve.  A high-order tetrahedral volume mesh of the cardiac geometry is then generated for the end-disatole image frame based on the high-order surface mesh.  This is done using our advancing front approach, which is described in [2] and extends our high-order traingular mesh genneration technique in [3].  We employ the developed deformable image registration technique to warp the surface meses of the left and right ventricles from end-diastole to the remaining cardiac phases, hence propagating the end-diastole surface meshes throughout the cardiac cycle.  We then apply a high-order finite element approach to deform the volume mesh at end-diastole throughout the cardiac cycle based on the warped surface meshes at the corresponding cardiac phases.  We compare the volume meshes generated via our finite element-based mesh warping approach with the corresponding volume meshes obtained directly by warping the end-diastole volume mesh using the displacement field yielded by the deformable image registration method.  We assess the performance of these methods on the Automated Cardiac Diagnosis Challenge Dataset.

REFERENCES

[1]   Roshan R. Upendra, Brian Wentz, Suzanne M. Shontz, and Cristian A. Linte, A convolutional neural network-based deformable image registration method for cardiac motion estimation from cine cardiac MR images, Computing in Cardiology, IEEE, 47:1-4, 2020

[2]  Fariba Mohammadi and Suznane M. Shontz (2021, February). A Direct Method of Generating High-order Tetrahedral Meshes Using an Advancing Front Approach, Submitted to the 29^th International Meshing Roundtable.

[3]   Fariba Mohammadi, Shusil Dangi, Suzanne M. Shontz, & Cristian A. Linte (2020, June). A Direct High-Order Curvilinear Triangular Mesh Generation Method Using an Advancing Front Technique. In Proceedings of the 2020 International Conference on Computational Science (pp. 72-85). Springer, Cham.

Presenting Author: Suzanne Shontz University of Kansas

Authors:

Fariba Mohammadi University of Kansas
Brian Wentz University of Kansas
Roshan Upendra Rochester Institute of Technology
Suzanne Shontz University of Kansas
Cristian Linte Rochester Institute of Technology