Clinical Metadata-Guided Limited-Angle CT Image Reconstruction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Science

Background:

• Limited-angle computed tomography (LACT) offers benefits like reduced radiation dose and improved temporal resolution.
• Severe artifacts arise in LACT due to missing projection data, hindering clinical application.
• Existing reconstruction methods underutilize valuable patient- and acquisition-level metadata.

Purpose of the Study:

• To develop a robust image reconstruction framework for LACT by integrating clinical metadata as semantic priors.
• To address the ill-posed nature of LACT reconstruction using a novel two-stage diffusion model.
• To enhance anatomical fidelity and quantitative accuracy in LACT images.

Main Methods:

• A two-stage conditional diffusion model framework was proposed, leveraging clinical metadata (e.g., acquisition parameters, patient demographics).
• Stage-I: A transformer-based diffusion model generates a coarse anatomical estimate conditioned on metadata.
• Stage-II: A second diffusion model refines the estimate, incorporating metadata and physics-based data consistency (ADMM).

Main Results:

• The metadata-guided diffusion framework significantly outperformed iterative, CNN-based, and metadata-free diffusion baselines on the CTRATE dataset.
• Substantial improvements in image quality (SSIM/PSNR) were observed under severe angular truncation (up to 90°).
• Clinical validation on cardiac CT showed improved coronary artery calcium scoring accuracy compared to full-view references.

Conclusions:

• Clinical metadata can be effectively utilized as semantic priors to enhance LACT image reconstruction.
• The proposed framework demonstrates superior performance in fidelity and quantitative accuracy, especially under challenging limited-angle conditions.
• This approach holds significant potential for improving clinical utility and diagnostic confidence in LACT applications.