Accelerating cardiovascular model building with convolutional neural networks.

Area of Science:

• Biomedical Engineering
• Medical Imaging
• Artificial Intelligence

Background:

• Patient-specific cardiovascular models are crucial for simulations but require extensive manual 2D segmentation.
• Current segmentation methods are time-consuming and labor-intensive, hindering efficient model construction.

Purpose of the Study:

• To develop and validate a fully convolutional neural network (FCNN) method to automate and improve 2D cardiovascular segmentation.
• To reduce user effort and turnaround time for building patient-specific cardiovascular models using SimVascular software.

Main Methods:

• A fully convolutional neural network (FCNN) was trained using a novel loss function for partially labeled data to generate 2D vessel enhancement images.
• The marching-squares algorithm extracted segmentations from FCNN-generated images, which were then used for 3D model construction.
• An automated quality control method was implemented to select accurate segmentations.

Main Results:

• The FCNN method demonstrated improved 2D segmentation accuracy compared to traditional threshold and level set algorithms.
• Automated quality control enhanced the average DICE score by 25.8% and led to user acceptance of 80% of segmentations.
• The FCNN approach reduced manual segmentation effort by up to 73%, decreasing cardiovascular simulation turnaround time.

Conclusions:

• The proposed FCNN-based method effectively reduces manual segmentation effort for patient-specific cardiovascular model building.
• This automated approach improves segmentation accuracy and efficiency, facilitating faster cardiovascular simulations.
• The method is applicable to general tubular structure segmentation beyond cardiovascular applications.