LinFlo-Net: A Two-Stage Deep Learning Method to Generate Simulation Ready Meshes of the Heart.

Area of Science:

• Medical imaging
• Computational modeling
• Artificial intelligence

Background:

• Generating accurate computational models of the human heart from patient imaging data is crucial for clinical applications.
• Existing deep learning methods for cardiac mesh generation can suffer from self-penetration issues, requiring postprocessing.

Purpose of the Study:

• To develop a deep learning model for automatic generation of patient-specific cardiac computer models.
• To emphasize the generation of thin-walled cardiac structures free from mesh self-intersections.

Main Methods:

• A two-stage diffeomorphic deformation process is employed to fit a template mesh to cardiac imaging data.
• A novel loss function, derived from kinematics, penalizes surface contact and interpenetration during mesh deformation.

Main Results:

• The model achieves accuracy comparable to state-of-the-art methods.
• The generated cardiac meshes are free of self-intersections, unlike those from some prior methods.
• Resultant meshes are directly usable in physics-based simulations.

Conclusions:

• The proposed deep learning framework effectively generates accurate and topologically sound cardiac models from patient imaging.
• This method reduces the need for manual postprocessing, streamlining the workflow for cardiac simulations.