Creating a Digital Twin of the Human Heart Using Machine Learning
Disciplines
Biomechanical Engineering | Mechanical Engineering
Abstract (300 words maximum)
Personalized simulations of the left atrium enable professionals to create patient-specific models that replicate the electrophysiological and biomechanical behavior of the atrium. Patient-specific CT/MRI data is segmented to create finite element analysis (FEA) meshes, which enable precise modeling. However, manual segmentation is time-consuming and prone to errors. Additionally, existing neural network-based tools fail to capture detailed anatomical features limiting their use for FEA simulations and personalized treatment planning. This research aims to develop an automated neural network-based tool to accurately segment the left atrium and create a ready-to-use FEA mesh for left atrium simulations, or a digital twin of the human heart. To address these challenges, the focus was on the intricate anatomy of the left atrium, including the appendage and trabeculations, to improve both segmentation accuracy and simulation quality. To achieve this, a neural network will be trained using 120 3D CT/MRI images from public datasets and 100 CT images from Emory University collaborators. The images are then manually segmented to render a replica of the left atrium through the SimVascular workflow. The segmentation is then smoothed using the software tool Meshmixer to create high-quality training datasets. Data augmentation will enhance the dataset. A HeartDeformNet-derived neural network will be implemented for segmentation, using 150 images for training, 20 for validation, and 50 for testing. The predicted meshes will be compared with manually segmented ones using the DICE score and other quality metrics. Based on the quality metrics, necessary adjustments will be made to the results. The predicted FEA mesh will be integrated with a 0D lumped-parameter model to simulate full-cycle left atrium function based on patient-specific inputs such as ECG signals and pressure measurements. This tool is expected to reduce segmentation time and errors while improving simulation accuracy, contributing to better pre-surgical planning and personalized cardiovascular treatments.
Academic department under which the project should be listed
SPCEET - Mechanical Engineering
Primary Investigator (PI) Name
Lei Shi
PDF Version of Presentation
Creating a Digital Twin of the Human Heart Using Machine Learning
Personalized simulations of the left atrium enable professionals to create patient-specific models that replicate the electrophysiological and biomechanical behavior of the atrium. Patient-specific CT/MRI data is segmented to create finite element analysis (FEA) meshes, which enable precise modeling. However, manual segmentation is time-consuming and prone to errors. Additionally, existing neural network-based tools fail to capture detailed anatomical features limiting their use for FEA simulations and personalized treatment planning. This research aims to develop an automated neural network-based tool to accurately segment the left atrium and create a ready-to-use FEA mesh for left atrium simulations, or a digital twin of the human heart. To address these challenges, the focus was on the intricate anatomy of the left atrium, including the appendage and trabeculations, to improve both segmentation accuracy and simulation quality. To achieve this, a neural network will be trained using 120 3D CT/MRI images from public datasets and 100 CT images from Emory University collaborators. The images are then manually segmented to render a replica of the left atrium through the SimVascular workflow. The segmentation is then smoothed using the software tool Meshmixer to create high-quality training datasets. Data augmentation will enhance the dataset. A HeartDeformNet-derived neural network will be implemented for segmentation, using 150 images for training, 20 for validation, and 50 for testing. The predicted meshes will be compared with manually segmented ones using the DICE score and other quality metrics. Based on the quality metrics, necessary adjustments will be made to the results. The predicted FEA mesh will be integrated with a 0D lumped-parameter model to simulate full-cycle left atrium function based on patient-specific inputs such as ECG signals and pressure measurements. This tool is expected to reduce segmentation time and errors while improving simulation accuracy, contributing to better pre-surgical planning and personalized cardiovascular treatments.