Session: 06-11-02: Biotechnology and General Applications

Paper Number: 150820

150820 - Automated, Open-Source Workflow for Creation of Human Digital Twins at Scale 

Introduction

In concert with experiment and testing, computational modeling has helped to advance the understanding of Traumatic Brain Injury (TBI) due to its ability to simulate injury-producing environments that may not be practical or ethical in an experimental context.  Many state-of-the-art approaches rely on a single finite element head model due, in part, to the large investment of resources required to make a de novo high-fidelity model.  A single model lacks the ability to provide insights for populations. To address this singleton model shortcoming, class-based (e.g., male/female or small/medium/large) and mesh morphing strategies have been proposed.  

All the foregoing approaches use manual (human-in-the-loop) methods for medical image segmentation and finite element mesh generation.  Manual segmentation and mesh generation impose significant bottlenecks in the process of model generation, prohibiting large, population-based analyses to be undertaken.

Moreover, class-based and morph-based approaches can fail to faithfully reproduce the actual anatomy of any specific patient, potentially preventing patient-specific injury risk assessment to be made. 

The lack of scalability and specificity are significant impediments toward the goal of personalized injury risk assessment for large populations.  An automated system that uses patient-specific medical images to create unique, personalized digital twin models would help to alleviate these shortcomings.  The goal of this work is to demonstrate a fully automated, open-source workflow to create a large population (100+) of patient-specific head models based on each patient’s unique medical imaging.

Methodology

We utilized two open access repositories of human head MRIs, the IXI and SCI datasets, as raw input start points.   From these datasets, over 600 patient records were available.  We randomly down-selected 100 records and performed unique image segmentation and mesh generation in a fully automated manner. 

For brain MRI segmentation, we used SynthStrip without cerebral spinal fluid (CSF) for brain segmentation (freesurfer version 7.4.1), SynthSeg for brain segmentation (freesurfer version 7.4.1), BET for skull stripping (fsl version 6.0.7.8) followed by FAST for brain segmentation (fsl version 6.0.7.8), and SynthStrip for skull stripping (freesurfer version 7.4.1) followed by FAST for brain segmentation (fsl version 6.0.7.8).

We then combined the results from these four separate software workflows in an ensemble via hard voting, where each voxel was classified as either “brain tissue” or “not brain tissue.”  The interstitial voxels between the brain and skull were categorized as cerebral spinal fluid. 

The automated workflow produced a semantic segmentation, indicating the geometric structure of skull, CSF, and brain.  The semantic segmentation was then used to construct a high-fidelity finite element mesh, with a one-to-one correspondence from a classified image voxel to hexahedral finite element of the same material.   Auto-generated finite element meshes were sole-substituted for manually generated finite element meshes previously used in simulations with experimentally validated results.  Strain and strain rate histories generated from the automated and manual approaches were compared.

Results

One hundred patient-specific finite element models were produced without human intervention.  Models typically resulted in about one-quarter million bone elements, about one-half million CSF elements, and over one million brain elements.   When used in finite element simulations, automatically generated models used in place of previously validated manual models produced consistent, yet unique, strain and strain rate histories.  Model-to-model comparison showed high variability in the magnitude of the peak strain and peak strain rates.  Localization of the specific gyri of the cerebral cortex where the peak strain occurred also varied across models.  These results highlight the importance of accurately modeling the subject-specific cortical folds in a human digital to predict the location of injury. Our automated workflow captured these unique anatomical details.

Conclusions

We have developed and demonstrated a fully automated and open-source workflow for the creation of subject-specific high-quality finite element models at scale, directly from the medical images. This workflow establishes the groundwork for high-fidelity real-time personalized TBI risk assessment.

Presenting Author: Anu Tripathi Robert Morris University

Presenting Author Biography: Dr. Tripathi has expertise in the creation of high-fidelity finite element models of the human
head for the purposes of injury risk assessment and mitigation. She currently holds a post-
doctoral position with the Carlsen laboratory at Robert Morris University, and is actively
collaborating on research with Boston University, Sandia National Laboratories, and University
of Wisconsin Madison.

Authors:

Anu Tripathi Robert Morris University
Michael Buche Sandia National Laboratories
Rika Carlsen Robert Morris University
Chad Hovey Sandia National Laboratories
Emma Lejeune Boston University