LinFlo-Net: A two-stage deep learning method to generate simulation ready meshes of the heart

TL;DR

LinFlo-Net tackles the problem of generating simulation-ready, thin-walled whole-heart meshes from patient imaging by a two-stage template deformation: a global linear transformation followed by a diffeomorphic flow learned from images. A novel unsigned-distance input, CFL-inspired flow clipping, and a physics-based volume loss jointly ensure accurate, bijective deformations that avoid self-intersections while preserving tissue thickness where visible. Empirical results show competitive accuracy with state-of-the-art methods and, crucially, near-elimination of self-intersections, yielding meshes readily usable for physics-based heart simulations with minimal post-processing. The method demonstrates strong generalization to new templates and provides a practical path toward automated, high-quality patient-specific cardiac mesh generation for biomechanical analysis and simulation.

Abstract

We present a deep learning model to automatically generate computer models of the human heart from patient imaging data with an emphasis on its capability to generate thin-walled cardiac structures. Our method works by deforming a template mesh to fit the cardiac structures to the given image. Compared with prior deep learning methods that adopted this approach, our framework is designed to minimize mesh self-penetration, which typically arises when deforming surface meshes separated by small distances. We achieve this by using a two-stage diffeomorphic deformation process along with a novel loss function derived from the kinematics of motion that penalizes surface contact and interpenetration. Our model demonstrates comparable accuracy with state-of-the-art methods while additionally producing meshes free of self-intersections. The resultant meshes are readily usable in physics based simulation, minimizing the need for post-processing and cleanup.

Figures (12)

• Figure 1: Illustrative examples of CT (top row) and MRI (bottom row) images of the cardiac region. The segmentation of some cardiac structures of interest are overlaid on the figures on the right. Since the myocardium is a thick muscular structure, it is clearly visible in the image. However, the tissue thickness of other structures like the aorta are not visible in these images, and only the blood pool within these structures is visible.
• Figure 2: Two samples from our dataset that demonstrate the variation in scale across samples. Black wireframe is the template mesh, and red surface is the ground-truth mesh.
• Figure 3: Workflow describing the training pipeline for the linear transformation module
• Figure 4: Workflow illustrating the training pipeline for the flow deformation module
• Figure 5: An illustration of samples generated by the data-augmentation process from a given input image and associated ground-truth segmentation. We generate these samples by applying small perturbations including random scaling, translation, rotation, shear, and local b-spline deformations to the input image and segmentation. Meshes are produced from the generated segmentations via the marching cubes algorithm.