Bone-GAN: Generation of virtual bone microstructure of high resolution peripheral quantitative computed tomography

Area of Science:

• Biomedical Engineering
• Radiology
• Computational Biology

Background:

• Developing medical biomarkers for bone health requires extensive image data, which is often limited by the size and quality of physical measurements.
• Current datasets are insufficient for robust training of data-driven models for bone analysis.

Purpose of the Study:

• To develop a reliable in silico method for generating realistic synthetic bone microstructures with controllable properties.
• To enhance the training of neural networks for bone analysis and facilitate the development of new diagnostic parameters for bone architecture and mineralization.

Main Methods:

• A progressive volumetric generative adversarial network (GAN) was trained on 10,795 bone patches from human lumbar vertebrae (high-resolution peripheral quantitative CT).
• A style transfer technique was integrated to enable the generation of synthetic samples with specific microarchitectural and gestalt properties by optimizing entangled loss functions.
• Reliability was assessed by comparing 10 microstructural parameters between real and synthetic bone samples.

Main Results:

• The developed method successfully generated synthetic bone samples that closely matched real samples in both visual and quantitative aspects.
• The GAN's latent space allowed for smooth morphing of bone samples based on microstructural parameters and visual appearance.
• Optimal synthesis was achieved for 32x32x32 voxel samples, with successful generation of 64x64x64 voxel samples.

Conclusions:

• A two-step approach combining a parameter-agnostic GAN with parameter-specific style transfer generates unlimited, anonymous, and realistic microstructural bone data.
• This synthetic database supports the development of novel data-driven bone biomarker methods.
• The style transfer capability allows for the simulation of specific bone pathologies by generating datasets under defined conditions.