When are Diffusion Priors Helpful in Sparse Reconstruction? A Study with Sparse-view CT.

Area of Science:

• Medical imaging
• Artificial intelligence in healthcare
• Computational imaging

Background:

• Diffusion models achieve state-of-the-art image generation performance.
• Their application in sparse medical image reconstruction is emerging.
• Unlike classical methods, diffusion models can generate realistic images even when inaccurate, especially with limited data.

Purpose of the Study:

• To investigate the effectiveness of diffusion models as priors for sparse medical image reconstruction.
• To compare diffusion model performance against classical priors (sparse and Tikhonov regularization).
• To evaluate performance across varying numbers of observations using pixel-based, structural, and downstream metrics on low-dose chest wall CT for fat mass quantification.

Main Methods:

• Comparative analysis of reconstruction algorithms.
• Varying the number of projection data points.
• Utilizing pixel-based, structural, and downstream evaluation metrics.
• Application to low-dose chest wall CT imaging for fat mass quantification.

Main Results:

• Classical priors outperform diffusion priors when a sufficient number of projections are available.
• Diffusion priors capture significant detail with very few observations, surpassing classical methods.
• Diffusion model performance plateaus after approximately 10-15 projections, failing to capture all details even with more data.

Conclusions:

• Diffusion models offer advantages in low-data regimes for sparse medical image reconstruction.
• Classical priors remain superior in scenarios with adequate data.
• Potential pitfalls of diffusion-based reconstruction necessitate further research, especially for clinical applications.