AI-Powered Automated Model Construction for Patient-Specific CFD Simulations of Aortic Flows (2503.12515v1)

Abstract: Image-based modeling is essential for understanding cardiovascular hemodynamics and advancing the diagnosis and treatment of cardiovascular diseases. Constructing patient-specific vascular models remains labor-intensive, error-prone, and time-consuming, limiting their clinical applications. This study introduces a deep-learning framework that automates the creation of simulation-ready vascular models from medical images. The framework integrates a segmentation module for accurate voxel-based vessel delineation with a surface deformation module that performs anatomically consistent and unsupervised surface refinements guided by medical image data. By unifying voxel segmentation and surface deformation into a single cohesive pipeline, the framework addresses key limitations of existing methods, enhancing geometric accuracy and computational efficiency. Evaluated on publicly available datasets, the proposed approach demonstrates state-of-the-art performance in segmentation and mesh quality while significantly reducing manual effort and processing time. This work advances the scalability and reliability of image-based computational modeling, facilitating broader applications in clinical and research settings.

• The paper introduces an integrated deep learning pipeline that automates constructing patient-specific aortic CFD meshes, significantly reducing manual errors and processing time.
• It leverages a novel voxel-based segmentation module with Bayesian uncertainty quantification and a GNN-LDDMM surface deformation module to generate smooth, accurate aortic models.
• Results demonstrate improved segmentation accuracy and CFD predictions, paving the way for efficient clinical and research applications in cardiovascular hemodynamics.

This paper, "AI-Powered Automated Model Construction for Patient-Specific CFD Simulations of Aortic Flows" (2503.12515), presents a deep learning framework designed to automate the traditionally manual and time-consuming process of creating simulation-ready patient-specific vascular models from medical images. The framework aims to improve the efficiency, consistency, and accuracy of image-based computational fluid dynamics (CFD) simulations, particularly for aortic flows, which are critical for understanding cardiovascular hemodynamics.

The core challenge addressed is the laborious nature of current workflows (like those in SimVascular), which involve manual steps such as centerline delineation, 2D slice segmentation, and surface reconstruction using methods like NURBS and Boolean operations. These manual steps are prone to errors, operator variability, and can take hours or days per patient, hindering clinical translation and large-scale research.

The proposed framework integrates two main components into a unified pipeline:

1. A Voxel-based Segmentation Module (LoGB-Net): This module is responsible for accurately delineating the vascular structure in 3D medical images at the voxel level.
   • Architecture: It's based on a 3D U-Net architecture but incorporates novel elements.
   • LoG Stream: Uses hierarchical 3D Laplacian-of-Gaussian (LoG) filters with multiple scales ($\sigma$ values corresponding to kernel sizes like 3, 5, 7, 9, 11). This allows the model to effectively detect vessels of varying diameters, from the main aorta to smaller branches. The kernel parameters can be dynamically optimized during training.
   • Bayesian Framework: The LoG module's training is formulated within a Bayesian framework. This treats parameters as random variables and estimates their probability distributions, providing uncertainty quantification (UQ). This is particularly useful for dealing with noisy or blurred image boundaries. Variational inference with ELBO maximization and a Dice loss is used for training.
   • Balanced Gate: Addresses data imbalance issues during training caused by random cropping, ensuring that the model sees an equal number of samples containing the main aorta and smaller branches. This is achieved by classifying cropped cubes based on the percentage of predicted foreground voxels and sampling equally from "large vessel" and "small branch" groups.
   • Implementation Details: Trained on cropped $64 \times 64 \times 64$ volumes from datasets (VMR, AVT). Preprocessing includes resampling, clipping, normalization, and background suppression (for training). Data augmentation (flipping, rotation, cropping) is applied.
2. A Surface Deformation Module (GNN-LDDMM): This module takes the voxel-based segmentation output from LoGB-Net and refines it into a smooth, high-fidelity 3D surface mesh suitable for CFD.
   • Input Processing: The voxel segmentation is converted to an initial triangular surface mesh using Marching Cubes. This mesh is then resampled (e.g., using ACVD) and smoothed (using normal, edge, and Laplacian regularization) to improve vertex distribution and reduce blockiness.
   • GNN Component: A graph neural network (using Chebyshev convolutions) processes the surface mesh points and predicts a vector field (initial momentum) over the surface.
   • LDDMM Component: The Large Deformation Diffeomorphic Metric Mapping (LDDMM) algorithm uses the predicted momentum field to iteratively deform the surface. LDDMM ensures that the transformation is diffeomorphic (smooth, invertible, preserves topology), preventing self-intersections and unrealistic distortions common in basic deformation methods. The deformation is governed by Hamiltonian dynamics integrated over time.
   • Scaling Gate: Prevents severe distortion at vessel inlets and outlets. A scalar field is computed based on geodesic distance from inlet/outlet centers. Points at the boundaries are immobilized ($\alpha=0$), points on the main vessel wall are fully mobile ($\alpha=1$), and points in between have graded mobility using a truncated Gaussian function. This scalar field scales the predicted deformation vectors.
   • Training: Unsupervised training using the input medical image. The objective function minimizes a misalignment energy term (negative log sum of image gradients over the surface, encouraging alignment with boundaries) and standard surface regularization terms (normal, edge, Laplacian losses) to maintain mesh quality.

Workflow Summary: The framework starts with the raw medical image, applies the LoGB-Net for initial voxel segmentation, then uses the GNN-LDDMM module to deform an initial surface (derived from the voxel output) to align with the image features, resulting in a simulation-ready mesh. This is a significant departure from manual SimVascular workflows.

Key Results and Practical Implications:

• Segmentation Accuracy: LoGB-Net outperforms several state-of-the-art deep learning segmentation methods (including various U-Net variants, Attention-based models, and Transformers) on aortic segmentation, achieving higher Dice coefficients and lower surface distance metrics (ASD, Hausdorff), particularly for challenging multi-branch geometries. The ablation paper confirms the importance of the LoG stream, Bayesian formulation, and balanced gate for this superior performance.
• Surface Reconstruction Quality: The GNN-LDDMM module refines the blocky voxel output into smooth, anatomically consistent surfaces. Detailed comparisons show that automated surfaces align better with image gradients than manual surfaces, which often exhibit oversmoothing or interpolation artifacts. The deformation module successfully reduces misalignment energy relative to static manual surfaces. Ablation studies demonstrate the necessity of both LDDMM (for smoothness and integrity) and the scaling gate (to prevent boundary distortions).
• CFD Relevance: The paper shows that even subtle geometric differences between manual and automated surfaces can lead to notable differences in simulated hemodynamic quantities like pressure and wall shear stress (WSS). The automated framework's improved geometric fidelity results in CFD solutions that capture finer details and potentially more accurate biomarker values.
• Uncertainty Quantification: The Bayesian nature of LoGB-Net allows the framework to produce an ensemble of plausible segmentation outcomes and propagate this uncertainty through the deformation and CFD steps. This moves beyond single-point predictions and provides valuable information about the reliability of the results, especially in regions with lower image quality (SNR). The deformation process significantly reduces uncertainty in the main aorta, but it can persist in noisy branch regions.
• Efficiency: The automated pipeline significantly reduces processing time compared to manual workflows, which is a major step towards making image-based CFD practical for clinical applications.

Limitations:

• The LDDMM module may struggle with very large initial deformations resulting from poor initial segmentation.
• The training dataset, while using public repositories, could benefit from being larger and more diverse to ensure robustness across a wider range of patient anatomies and imaging conditions.
• Performance can still be limited by severe image noise or poor quality, particularly in smaller branch vessels.

Broader Implications: The automated and robust nature of this framework has the potential to integrate CFD into clinical workflows by making model construction faster, more consistent, and less reliant on expert manual effort. The inclusion of UQ enhances the trustworthiness of the simulation results. The modular design could facilitate extensions to other vascular territories or organs and different imaging modalities.