Overview

Schema Version

https://atlas.rsna.org/schemas/2025-11/model.json

Name

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

Link

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC10698592

Indexing

Keywords: diffusion probabilistic model, super-resolution, CT, trabecular bone, proximal femur, osteoporosis, finite element analysis, BV/TV, trabecular thickness, trabecular spacing, trabecular number

Content: CT, MK

RadLex: RID10632, RID10374, RID6106, RID49214

SNOMED: 64859006

Author(s)

Trevor J. Chan

Chamith S. Rajapakse

Organization(s)

University of Pennsylvania

Version

1.0

Contact

Corresponding author: Chamith S. Rajapakse, PhD, Departments of Radiology and Orthopedic Surgery, University of Pennsylvania, 3400 Spruce St, Philadelphia, PA 19104-6243

Funding

Supported by National Science Foundation (grant no. 2026906) and National Institutes of Health (grant nos. R01 AR068282 and R01 AR076392).

Ethical review

Retrospective, HIPAA-compliant study approved by the institutional review board.

Date

Updated: 2023-11-01

Published: 2023-09-20

Created: 2022-11-15

References

[1] Chan TJ; Rajapakse CS. "A Super-Resolution Diffusion Model for Recovering Bone Microstructure from CT Images". Radiology: Artificial Intelligence. 2023 Nov;5(6):e220251. 2023-09-20. doi:10.1148/ryai.220251. PMID: 38074790. PMCID: PMC10698592.

Model

Architecture

Conditional denoising diffusion probabilistic model (DDPM) adapted from SR3; U-Net encoder-decoder with ResNet backbone; iterative reverse noising with T=2000 steps; conditioned on low-resolution image for 3× super-resolution.

Clinical benefit

May enable accurate measurements of trabecular bone structure and stiffness at radiation doses comparable to current clinical CT, improving viability of CT for assessing bone health and osteoporosis risk.

Clinical workflow phase

Clinical decision support/quantitative image postprocessing for bone health assessment; potential use in opportunistic assessment from routine CT.

Degree of automation

Automated image super-resolution and quantitative extraction of trabecular metrics from CT images; user selects ROIs for analysis in study setup.

Indications for use

Research-stage method intended to reconstruct trabecular bone microstructure of the proximal femur from lower-resolution CT to support assessment of bone health/osteoporosis risk in adults.

Input

Low-resolution CT sections of proximal femur (e.g., ~0.72 mm in-plane, 85×85 pixels), optionally stacked for 3D reconstruction.

Instructions

Condition the DDPM (SR3-style) on the low-resolution image and iteratively denoise for T=2000 steps to obtain a 3× upsampled image (~0.24 mm). Quantify BV/TV, TbTh, TbSp, TbN (e.g., with BoneJ) and estimate axial stiffness via finite element analysis on trabecular ROIs.

Limitations

Small dataset (26 cadaveric femurs); no validation on clinical in vivo data; low-resolution inputs simulated via bicubic downsampling may not capture scanner/protocol variability; high-resolution in vivo ground truth unavailable; potential biases due to limited population diversity.

Output

Description: Super-resolved high-resolution CT-like images of trabecular bone with derived quantitative microstructural metrics and estimated mechanical stiffness.

Recommendation

Promising for research and potential clinical translation; requires larger-scale and clinical validation before diagnostic use.

Regulatory information

Comment: No regulatory submissions reported.

Authorization status: Not a regulated/cleared medical device; research study.

Reproducibility

Model details provided (SR3-based DDPM; training/validation/test splits; optimization objective); supplemental materials reference implementation details; data available from corresponding author upon request.

Use

Intended: Image reconstruction

Out-of-scope: Image reconstruction, Diagnosis

Excluded: Other

User

Intended: Radiologist, Subspecialist diagnostic radiologist, Researcher

Out-of-scope: Layperson

Excluded: Other