Computational framework for the generation of one‐dimensional vascular models accounting for uncertainty in networks extracted from medical images

Michelle A. Bartolo, Alyssa Taylor-LaPole, Darsh Gandhi, Alexandria Johnson, Yaqi Li, Emma Slack, Isaiah Stevens, Zachary G. Turner, Justin Weigand, Charles Puelz, Dirk Husmeier, Mette S. OlufsenThe Journal of Physiology 2023

TL;DR

This paper introduces a novel computational framework to generate patient-specific 1D vascular models from medical images, specifically addressing the critical issue of uncertainty introduced by image segmentation and data extraction. By employing statistical change point analysis and robust algorithms for junction node placement and radius determination, the framework generates more reliable hemodynamic predictions, crucial for understanding cardiovascular health and disease.

What Problem Does This Paper Address?

One-dimensional (1D) cardiovascular models are powerful tools for non-invasively predicting hemodynamics like blood flow, pressure, and wave reflection. These models are built using vascular networks extracted from medical images (like CT and MRI). However, a significant challenge is the inherent uncertainty in this extraction process. Image segmentation, which identifies the vessels, is prone to variability due to factors like image resolution, patient motion, and user expertise. This uncertainty in vessel dimensions (radius, length) and network structure directly impacts the accuracy of hemodynamic predictions. Existing methods for generating vascular trees from image-derived centerlines often struggle to account for this uncertainty, leading to potentially unreliable model outputs. This paper aims to bridge this gap by developing a framework that explicitly quantifies and accounts for geometric uncertainty in the extracted vascular networks.

Methods

The core of this work is a novel framework that refines the process of generating 1D vascular models from medical images. Key methodological contributions include:

  • Statistical Change Point Analysis for Radius Extraction: Instead of averaging all radii along a vessel (which can be inaccurate, especially near junctions), the authors use statistical change point detection algorithms (R/changepoint and R/segmented) to identify segments within each vessel that best represent its true radius. This allows for a more accurate and robust radius estimation, including its standard deviation.
  • Junction Node Adjustment: VMTK (Vascular Modeling Toolkit), a common tool for centerline extraction, can sometimes place junction nodes outside of physiological regions (ostia). The framework includes an algorithm to shift these junction nodes to the center of the ostium, reclassifying centerline segments and improving the accuracy of parent and daughter vessel geometry.
  • Network Pruning: To enable meaningful comparisons between networks generated from different segmentations, especially in the pulmonary vasculature where vessel counts can vary, a pruning procedure is used to standardize the number of vessels to a consistent value.
  • Fractal Tree Augmentation: For small vessels not visible in the images, the framework augments the network with self-similar fractal trees, ensuring a complete representation down to a minimum radius.
  • 1D Fluid Dynamics Model: The generated labeled trees are then used with a 1D fluid dynamics model that solves Navier-Stokes equations for large vessels and a linearized wave equation for small vessels, allowing for hemodynamic simulations.

Datasets

The study analyzes two types of medical imaging data:

  • A chest CT image from a healthy, 67-year-old female volunteer.
  • An MRA (Magnetic Resonance Angiography) image from a patient with a double outlet right ventricle (DORV).

For the pulmonary vasculature, five segmentations were performed on the CT image. For the aortic vasculature, ten segmentations were performed on the MRA image. These multiple segmentations of the same images are crucial for assessing segmentation uncertainty.

Key Insights and Ideas

  • Uncertainty is Pervasive: The raw data extracted from medical images, even with the same user and software, exhibits significant variability in vessel dimensions and connectivity. This highlights the critical need for methods that account for this uncertainty.
  • Radius is King (for Uncertainty): The study found that variations in vessel radius have the most significant impact on hemodynamic predictions, even more so than junction location or length variations. This underscores the importance of accurately determining vessel radii.
  • Change Points are Powerful: Statistical change point analysis proves to be an effective and automated way to identify representative segments for radius estimation, overcoming limitations of simple averaging and accounting for complex vessel structures.
  • Post-processing is Essential: The framework demonstrates that post-processing steps like junction node adjustment and radius extraction using change points are vital for generating high-fidelity patient-specific models from image-derived data.
  • Standardization Matters: Pruning networks to a consistent vessel count significantly reduces variations in hemodynamic predictions, making comparisons more reliable.

Key Results

  • Reduced Prediction Variability: After applying the framework's post-processing steps (junction adjustment, radius extraction, and network pruning), the variability in pressure and flow predictions across different segmentations was significantly reduced. For example, pressure ranges in pulmonary arteries decreased from over 15 mmHg to about 5 mmHg after pruning.
  • Impact of Radius Variation: Simulations showed that even small perturbations in vessel radius can lead to substantial changes in hemodynamics, with a 70 mmHg range in systolic pressure observed for a small perturbation in the LCC radius.
  • Robustness of Change Point Analysis: The change point analysis protocol was found to be robust, requiring minimal post-processing for network consistency checks (less than 1% of vessels).
  • Significant Differences in Raw Data: One-way ANOVA tests revealed statistically significant differences in vessel length and radii among segmentations for several major arteries in both pulmonary and aortic networks, confirming the inherent uncertainty in raw VMTK outputs.

Limitations and Future Work

  • Independent Radius Sampling: The current framework samples radii for each vessel independently, overlooking the interconnectedness of parent and daughter vessels. Future work will account for covariance to address this morphometric condition.
  • Limited Number of Segmentations: Due to the time-consuming nature of manual segmentation, a relatively small number of segmentations were used. Future studies plan to use automatic segmentation software to generate more data for testing.
  • Assumptions in Modeling: The framework relies on certain assumptions regarding vessel radius evolution and junction locations. While customizable, these assumptions might need adjustment for different vascular networks (e.g., renal, cerebral). Future work will investigate the impact of specific modeling assumptions.
  • Exponential Tapering Model: The use of an exponential function for tapering radii is a simplification that reduces computational complexity. Future work could explore more complex models.
  • Image Noise: The impact of image noise on radius extraction is acknowledged, particularly for small vessels, and is a factor that could be further investigated.

Summary updated

Abstract

One-dimensional (1D) cardiovascular models offer a non-invasive method to answer medical questions, including predictions of wave-reflection, shear stress, functional flow reserve, vascular resistance, and compliance. This model type can predict patient-specific outcomes by solving 1D fluid dynamics equations in geometric networks extracted from medical images. However, the inherent uncertainty in in-vivo imaging introduces variability in network size and vessel dimensions, affecting hemodynamic predictions. Understanding the influence of variation in image-derived properties is essential to assess the fidelity of model predictions. Numerous programs exist to render three-dimensional surfaces and construct vessel centerlines. Still, there is no exact way to generate vascular trees from the centerlines while accounting for uncertainty in data. This study introduces an innovative framework employing statistical change point analysis to generate labeled trees that encode vessel dimensions and their associated uncertainty from medical images. To test this framework, we explore the impact of uncertainty in 1D hemodynamic predictions in a systemic and pulmonary arterial network. Simulations explore hemodynamic variations resulting from changes in vessel dimensions and segmentation; the latter is achieved by analyzing multiple segmentations of the same images. Results demonstrate the importance of accurately defining vessel radii and lengths when generating high-fidelity patient-specific hemodynamics models.
BibTeX
@article{bartolo2023computational,
  author = {Michelle A. Bartolo and Alyssa Taylor-LaPole and Darsh Gandhi and Alexandria Johnson and Yaqi Li and Emma Slack and Isaiah Stevens and Zachary G. Turner and Justin Weigand and Charles Puelz and Dirk Husmeier and Mette S. Olufsen},
  title = {{Computational framework for the generation of one‐dimensional vascular models accounting for uncertainty in networks extracted from medical images}},
  journal = {The Journal of Physiology},
  year = {2023},
}
Figure