Related Experiment Video
Updated: Sep 21, 2025
07:45
Author Spotlight: Optimizing Dendritic Spine Analysis for Balanced Manual and Automated Assessment in the Hippocampus CA1 Apical Dendrites
Published on: September 27, 2024
2.4K
Generative Adversarial Network (GAN) for Automatic Reconstruction of the 3D Spine Structure by Using Simulated
Ching-Juei Yang1,2, Cheng-Li Lin3,4, Chien-Kuo Wang5
1Department of Biomedical Engineering, National Cheng Kung University, Tainan 701401, Taiwan.
Diagnostics (Basel, Switzerland)
|May 28, 2022
Summary
This study developed a 2D to 3D generative adversarial network (GAN) for automatic spine reconstruction from X-rays. The model shows potential for clinical use in generating 3D spine models from 2D images.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Spine Anatomy
Background:
- Generating 3D spine models from 2D X-rays is challenging.
- Existing methods may lack accuracy and efficiency.
Purpose of the Study:
- To develop and validate a 2D to 3D generative adversarial network (GAN) for automatic spine reconstruction.
- To assess the feasibility of producing 3D spine structures from simulated bi-planar 2D X-ray images.
Main Methods:
- Modified X2CT-GAN into a 2D to 3D GAN for spine imaging.
- Utilized retrospective CT data from 984 patients (1012 studies).
- Evaluated model performance using 10-fold cross-validation and metrics like DSC, JSC, OV, and SSIM.
Main Results:
- Optimal mean DSC: 0.8192, JSC: 0.6984, OV: 0.8624, SSIM_AP: 0.9261, SSIM_Lat: 0.9242.
- Training performance improved with larger datasets and enhanced bone signal conditions.
- Demonstrated significant improvement in training performance.
Conclusions:
- The 2D to 3D GAN shows potential for clinical implementation in automatic 3D spine image generation.
- This prototype can be a foundation for advanced medical diagnostic techniques using transfer learning.
- The model facilitates automatic production of 3D spine images from 2D X-rays.
