Related Experiment Video
Updated: Dec 20, 2025
13:34
A Combined 3D Tissue Engineered In Vitro/In Silico Lung Tumor Model for Predicting Drug Effectiveness in Specific Mutational Backgrounds
Published on: April 6, 2016
10.5K
An adversarial machine learning framework and biomechanical model-guided approach for computing 3D lung tissue
Anand P Santhanam1, Brad Stiehl1, Michael Lauria1
1Department of Radiation Oncology, University of California, Los Angeles, Los Angeles, CA, 90095, USA.
Medical Physics
|May 26, 2020
Summary
This study introduces a machine learning method using a constrained generalized adversarial neural network (cGAN) to predict lung tissue elasticity from CT scans. This approach enables accurate elasticity estimation without requiring four-dimensional (4D) imaging, crucial for radiotherapy and beyond.
Area of Science:
- Medical Imaging
- Computational Biomechanics
- Machine Learning
Background:
- Lung elastography measures tissue elasticity for diagnostics and biomechanical applications.
- Current methods require four-dimensional computed tomography (4DCT) for lung motion estimation.
- 4DCT is primarily used in radiotherapy, limiting broader applications of lung elastography.
Purpose of the Study:
- To develop a machine learning-based method for predicting lung tissue elasticity from a single end-expiratory 3D CT scan.
- To enable lung elasticity prediction in scenarios where 4D imaging is unavailable.
- To support applications in radiotherapy and other domains requiring lung biomechanical characterization.
Main Methods:
- A deep neural network, specifically a constrained generalized adversarial neural network (cGAN), was employed.
- The cGAN was trained using five-dimensional CT (5DCT) datasets and a finite element biomechanical lung model.
- Inverse elasticity estimation was performed by iteratively solving for elasticity to match ground-truth deformations, with the cGAN learning this relationship in a supervised manner.
Main Results:
- The cGAN approach effectively computed lung tissue elasticity from end-expiratory CT, achieving high accuracy on validation datasets (0.87 ± 0.4 KPa).
- Generated end-inhalation CT images using the estimated elasticity closely matched original CT scans, with high similarity metrics (MI: 1.77, SSIM: 0.89, NCC: 0.97).
- The method demonstrated near real-time computation of lung tissue elasticity.
Conclusions:
- A cGAN-based machine learning method can accurately predict lung tissue elasticity from standard end-expiratory CT images.
- This technique bypasses the need for 4DCT, expanding the applicability of lung elastography.
- The predicted elasticity, combined with a biomechanical model, allows for clinically acceptable generation of 4D lung deformations.
