A Framework for Guiding DDPM-Based Reconstruction of Damaged CT Projections Using Traditional Methods

Area of Science:

• Medical Imaging
• Computational Imaging
• Artificial Intelligence in Healthcare

Background:

• Denoising Diffusion Probabilistic Models (DDPM) show promise for image synthesis but struggle with detail preservation in computed tomography (CT) reconstruction.
• Compromised CT projection data (limited-angle, sparse-view, low-dose) present significant challenges for accurate image reconstruction.

Purpose of the Study:

• To develop and validate a novel hybrid framework (PHIRF) integrating conventional CT reconstruction with DDPM for enhanced image quality from compromised projection data.
• To improve the preservation of fine anatomical details and reduce artifacts in CT images reconstructed from ill-posed projection scenarios.

Main Methods:

• A dual-phase approach combining conventional CT reconstruction algorithms (FBP, ART, ML-EM) with DDPM.
• Conventional methods generate preliminary reconstructions and low-dimensional features from incomplete projections.
• These features condition the DDPM's reverse diffusion process for synthesizing enhanced tomographic images.

Main Results:

• The PHIRF framework achieved state-of-the-art performance across limited-angle, sparse-view, and low-dose CT projection scenarios.
• Demonstrated superior preservation of fine anatomical details compared to existing deep learning methods.
• Significantly suppressed reconstruction artifacts, outperforming current deep learning-based reconstruction approaches.

Conclusions:

• The proposed PHIRF framework effectively combines physical prior knowledge with data-driven generative models for medical image reconstruction.
• PHIRF offers a robust and adaptable solution for enhancing CT image quality from various forms of compromised projection data.
• This hybrid architecture represents a new paradigm for advanced CT image reconstruction, particularly in challenging clinical scenarios.