SimVascular

SimVascular constructs patient-specific cardiovascular models by automating 2D vessel segmentation and reconstructing 3D vessel geometries for computational simulations.

Key Features:

• Fully Convolutional Neural Network (FCNN): A fully convolutional neural network computes local vessel enhancement images with per-pixel likelihoods of vessel tissue.
• Novel loss function: The FCNN is trained with a loss function specifically tailored for partially labeled segmentation data.
• Pathline-based 2D extraction: Pixel intensities are extracted from images perpendicular to user-supplied vessel pathlines to form 2D input images.
• Marching-squares segmentation: The marching-squares algorithm is applied to FCNN-enhanced images to extract central vessel segmentations.
• 3D transformation and interpolation: Extracted 2D vessel surface points are transformed back into 3D and interpolated along pathlines to form final vessel models.
• Automated quality control: An automated quality-control method selects promising segmentations to improve overall segmentation accuracy and acceptance.
• Quantitative improvement: Demonstrated a 25.8% increase in average DICE score versus threshold and level set algorithms and achieved 80% user acceptance for the best-performing FCNN.
• Reduced manual effort: Reported reduction of manual segmentation effort by up to 73%, accelerating model-building turnaround.
• General tubular applicability: The method is adaptable to general tubular structures in addition to cardiovascular vessels.
• Software components: Includes components svModel, svMesh, svPre, and svPost for model building and processing.

Scientific Applications:

• Patient-specific cardiovascular modeling: Building individualized vascular geometries for simulation and analysis.
• Medical image vessel segmentation: Automating 2D vessel segmentation in medical imaging datasets.
• 3D tubular reconstruction: Reconstructing 3D geometries of tubular anatomical structures from 2D segmentations and pathlines.
• Segmentation evaluation and QC: Supporting segmentation accuracy assessment using metrics such as the DICE score and automated selection of high-quality segmentations.

Methodology:

Input image data and vessel pathlines; extract pixel intensities perpendicular to pathlines to create 2D images; process 2D images with an FCNN trained using a loss function for partially labeled data to produce vessel enhancement images; apply marching-squares to extract central vessel segmentations; transform 2D points back into 3D and interpolate along pathlines; apply automated quality control to select promising segmentations.