Deep learning-based harmonization of trabecular bone microstructures between high- and low-resolution CT imaging

Area of Science:

• Medical Imaging
• Biomedical Engineering
• Osteoporosis Research

Background:

• Osteoporosis diagnosis relies on bone mineral density and microstructure.
• Clinical CT enables in vivo bone microstructural imaging.
• Variations in CT scanner resolution necessitate image harmonization for consistent metrics.

Purpose of the Study:

• To develop and evaluate a deep learning (DL) method for harmonizing bone microstructural images from low- and high-resolution CT scanners.
• To assess the method's performance on image data and derived microstructural metrics.

Main Methods:

• A 3D version of GAN-CIRCLE, utilizing two generative adversarial networks (GANs), was developed for CT image resolution harmonization.
• The model learned to map low-resolution CT (LRCT) to high-resolution CT (HRCT) and vice versa.
• Supervised and unsupervised training/evaluation were performed on LRCT and HRCT image blocks from 20 volunteers.

Main Results:

• Supervised and unsupervised 3DGAN-CIRCLE methods significantly improved structural similarity (SSIM) compared to LRCT.
• Supervised 3DGAN-CIRCLE demonstrated higher agreement (CCC) for trabecular (Tb) measures compared to LRCT and unsupervised methods.
• The supervised method reduced bias and variability in Tb measures, outperforming existing DL methods.

Conclusions:

• 3DGAN-CIRCLE effectively generates HRCT images with high structural similarity to true HRCT.
• Supervised 3DGAN-CIRCLE enhances the accuracy of microstructural measures and outperforms unsupervised approaches.
• This DL solution aids in harmonizing multi-site imaging data for longitudinal osteoporosis studies.