CT Reconstruction using Nonlinear Diffusion Posterior Sampling with Detector Blur Modeling.

Area of Science:

• Medical Imaging
• Computational Imaging
• Deep Learning

Background:

• Deep learning has improved CT image quality via denoising.
• Spatial resolution, not noise, often limits CT applications and diagnostics.
• Existing methods struggle with enhancing spatial resolution in CT reconstructions.

Purpose of the Study:

• To improve spatial resolution in CT reconstructions.
• To combine deep learning with physical modeling of detector blur.
• To leverage diffusion models as deep image priors for CT deblurring and reconstruction.

Main Methods:

• Utilized diffusion models as deep image priors for regularizing CT reconstruction.
• Employed Diffusion Posterior Sampling (DPS) to integrate a deep prior with a measurement likelihood.
• Developed a nonlinear model accounting for detector blur post-Beer-Lambert attenuation.
• Trained a score estimator for a CT score-based prior and applied Bayes rule.

Main Results:

• Demonstrated the approach using simulated CT data.
• Compared the Diffusion Posterior Sampling (DPS) approach with filtered-backprojection (FBP) and model-based iterative reconstruction (MBIR).
• Observed a particular advantage for DPS in low-exposure CT data.
• Reported significant differences in error profiles between DPS and classical methods.

Conclusions:

• The proposed method effectively improves spatial resolution in CT reconstructions.
• Diffusion Posterior Sampling (DPS) shows promise for enhancing low-exposure CT imaging.
• This approach offers a novel way to combine deep priors with physical models for CT deblurring.