A Super-Resolution Diffusion Model for Recovering Bone Microstructure from CT Images.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Osteoporotic fractures commonly occur in the proximal femur.
• Assessing bone microstructure is crucial for predicting fracture risk.
• Current imaging methods may have limitations in resolving fine bone structures.

Purpose of the Study:

• To develop a diffusion-based deep learning model for reconstructing high-resolution bone microstructure from low-resolution CT images.
• To evaluate the model's accuracy in visualizing and quantifying trabecular bone structure in the proximal femur.

Main Methods:

• A retrospective study using 26 cadaveric micro-CT scans as ground truth.
• Downsampling high-resolution scans to create low-resolution training data.
• Employing a diffusion-based deep learning model to increase image resolution threefold.
• Validating model performance using microstructural metrics and finite element analysis.

Main Results:

• The proposed model demonstrated superior accuracy (ICC 0.92) compared to baseline methods (ICC 0.83).
• The model exhibited lower bias (3.80%) in physiologic metrics.
• Image quality metrics strongly correlated with accuracy (r > 0.89).

Conclusions:

• The diffusion-based deep learning model accurately recovers bone microstructure from low-resolution CT.
• This approach may enable precise bone structure and strength assessment using clinical CT protocols.
• The method holds potential for improving bone health evaluation and fracture risk prediction.