Predicting rat lumbar vertebral failure patterns as synthetic μCT images using a deep convolutional generative
Allison Tolgyesi1, Cari Whyne2, Michael Hardisty3
1Orthopaedic Biomechanics Laboratory, Sunnybrook Research Institute, 2075 Bayview Avenue, Toronto, ON, M4N 3M5, Canada; Institute of Biomedical Engineering, Faculty of Engineering, University of Toronto, 164 College Street, Toronto, ON, M5S 3G9, Canada.
Summary
This study developed a 3D conditional generative adversarial network (cGAN) to create synthetic micro-CT images of rat lumbar vertebrae fractures. The model accurately predicts bone failure patterns, showing promise for simulating biological damage.
Area of Science:
- Biomedical Engineering
- Medical Imaging
- Artificial Intelligence
Background:
- Predicting bone fracture patterns is crucial for understanding skeletal diseases and injury mechanisms.
- Micro-computed tomography (μCT) provides high-resolution 3D imaging of bone structure, but generating synthetic fracture images is challenging.
- Deep learning (DL) offers potential for creating realistic synthetic medical images.
Purpose of the Study:
- To develop and validate a 3D conditional generative adversarial network (cGAN) for predicting fracture patterns in rat lumbar vertebrae.
- To generate synthetic 3D μCT images of fractured vertebrae using DL.
- To assess the model's ability to simulate bone failure under axial compressive loading.
Main Methods:
- A 3D cGAN model was developed using sequential 3D μCT images of rat lumbar vertebrae under axial loading.
- Three predictive experiments were conducted: unloaded to 1500 μm displacement, 1500 μm to fracture, and unloaded to fracture.
- Model performance was evaluated using quantitative metrics (DSC, Jaccard, FID, SSIM) and qualitative assessment of fracture location and severity.
Main Results:
- The unloaded to 1500 μm experiment produced realistic loaded vertebrae images, maintaining metastatic disease presence and achieving the best quantitative metrics.
- The 1500 μm to fracture experiment outperformed the unloaded to fracture experiment in FID and SSIM metrics.
- The 1500 μm to fracture experiment showed improved fracture prediction accuracy (more true positives, fewer false negatives) with a low false positive rate (<10%).
Conclusions:
- The developed cGAN successfully generates realistic 3D μCT images of rat lumbar vertebrae failure patterns.
- The model demonstrates significant promise for future generative DL applications in modeling the damage behavior of biological structures.
- Predictive modeling of bone fractures using synthetic imaging can advance research in skeletal biomechanics and disease progression.
