Solving Zero-Shot Sparse-View CT Reconstruction With Variational Score Solver.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Computed tomography (CT) is essential for medical diagnosis but involves radiation exposure concerns.
• Reducing radiation dose in CT compromises image quality and diagnostic accuracy.
• Existing deep learning methods for CT reconstruction often require paired data, which is difficult to obtain.

Purpose of the Study:

• To develop a novel deep learning approach for sparse-view CT reconstruction without paired data.
• To address the challenge of maintaining image fidelity and diagnostic accuracy while reducing radiation dose in CT scans.

Main Methods:

• Introduced the Variational Score Solver (VSS), a method utilizing a latent diffusion model for sparse-view CT reconstruction.
• Acquired a probability distribution from densely sampled CT reconstructions using a diffusion model.
• Employed an iterative process integrating the diffusion model as a prior with the data consistency term.
• Distilled prior knowledge by finding the fixed point of the diffusion model, enabling precise control and a distribution-based approach to reconstruction.

Main Results:

• VSS demonstrated superior performance in sparse-view CT reconstruction compared to contemporary unsupervised methods.
• The method achieved results comparable to advanced supervised methods.
• Qualitative and quantitative experiments confirmed the effectiveness of VSS in producing high-quality reconstructions.

Conclusions:

• VSS offers a novel, unsupervised approach to zero-shot CT reconstruction, overcoming the limitations of supervised learning.
• The method effectively mitigates radiation concerns by enabling high-quality reconstructions from sparse-view data.
• VSS represents a significant advancement in low-dose CT imaging and diagnostic accuracy.