A Conditional Generative Diffusion Model of Trabecular Bone with Tunable Microstructure

Area of Science:

• Biomedical Engineering
• Materials Science
• Computational Biology

Background:

• Trabecular bone microstructure significantly influences skeletal integrity and fracture risk.
• Accurate digital models of trabecular bone are crucial for research, simulation, and developing new diagnostic tools.
• Existing methods for generating synthetic bone may lack precise control over key structural parameters.

Purpose of the Study:

• To develop a generative model capable of producing synthetic trabecular bone with tunable microstructural characteristics.
• To enable precise control over bone volume fraction (BV/TV), trabecular thickness (Tb.Th), and spacing (Tb.Sp) in generated bone samples.

Main Methods:

• Utilized Diffusion Transformers (DiT), a latent diffusion model with a transformer-based denoising network.
• Conditioned the model on target BV/TV, Tb.Th, and Tb.Sp values to guide microstructure generation.
• Trained the model on 29,898 ROIs from micro-CT scans of human femoral bone, with a 9:1 train/validation split.

Main Results:

• The model generated synthetic bone with good coverage of real-world microstructural distributions and visual realism.
• Achieved high Pearson correlations for target metrics: 0.9540 for BV/TV, 0.9618 for Tb.Th, and 0.9835 for Tb.Sp.
• Demonstrated stable microstructural characteristics across multiple model realizations, with low Coefficients of Variation (CV).

Conclusions:

• The developed generative model successfully creates realistic, tunable digital trabecular bone.
• This technology offers potential applications in virtual clinical trials and establishing priors for advanced image reconstruction.
• Enables precise control over synthetic bone architecture for diverse research applications.