Geometric Uncertainty in Patient-Specific Cardiovascular Modeling with Convolutional Dropout Networks

Area of Science:

• Biomedical Engineering
• Computational Fluid Dynamics
• Medical Imaging Analysis

Background:

• Accurate patient-specific cardiovascular models are crucial for understanding hemodynamics.
• Geometric uncertainty in imaging data can significantly influence model predictions.
• Existing methods often struggle to quantify or learn this geometric uncertainty directly.

Purpose of the Study:

• To develop a novel approach for generating patient-specific cardiovascular models from clinical image volumes.
• To integrate a convolutional neural network (CNN) for vessel lumen segmentation into a modeling pipeline.
• To quantify the impact of learned geometric uncertainty on hemodynamic parameters.

Main Methods:

• A CNN with dropout layers was trained for vessel lumen segmentation using a regression approach.
• Bayesian estimation was employed to derive vessel lumen surfaces from segmentation.
• The segmentation network was integrated into a path-planning pipeline to generate model families, enabling uncertainty quantification.

Main Results:

• The approach successfully generated patient-specific cardiovascular models incorporating learned geometric uncertainty.
• Geometric uncertainty was found to significantly impact wall shear stress and velocity magnitude, with coefficients of variation comparable to or exceeding other uncertainty sources.
• The impact of geometric uncertainty on pressure was limited, particularly in smaller vessels and rare lesion types.

Conclusions:

• The proposed method effectively learns and incorporates geometric uncertainty from training data into cardiovascular models.
• Geometric uncertainty is a critical factor influencing hemodynamic predictions, especially for shear stress and velocity.
• This approach offers a more robust framework for patient-specific cardiovascular modeling and uncertainty assessment.