Fast Computation of Hemodynamic Sensitivity to Lumen Segmentation Uncertainty

Area of Science:

• Biomedical Engineering
• Computational Fluid Dynamics
• Medical Imaging

Background:

• Patient-specific blood flow modeling aids coronary artery disease assessment.
• Accurate coronary segmentation and boundary conditions are crucial for diagnostic performance.
• Human review of segmentation is time-consuming, especially in non-critical regions.

Purpose of the Study:

• To develop a real-time machine learning framework for estimating geometric sensitivity in coronary artery models.
• To accelerate the review process of coronary segmentation by prioritizing high-impact regions.
• To improve the diagnostic accuracy of computational fluid dynamics simulations.

Main Methods:

• Developed a machine learning framework using geometric, clinical, and reduced-order model features.
• Utilized an anisotropic kernel regression for lumen narrowing score assessment.
• Introduced a multi-resolution sensitivity algorithm for hierarchical refinement of sensitive regions.

Main Results:

• The machine learning algorithm achieved a mean absolute error of less than 0.01 compared to 3D simulations.
• Demonstrated that sensitivity prediction incorporates hemodynamic information beyond simple anatomic reduction.
• Showcased the ability to quantify sensitivities to a desired spatial resolution.

Conclusions:

• The developed machine learning approach accurately estimates geometric sensitivity in real time.
• Focusing review on high-sensitivity regions significantly accelerates the simulation workflow.
• This sensitivity-driven approach has potential applications beyond coronary flow, including cerebral and electro-mechanical simulations.