Evaluation of Fluence Reduction versus Sparsity for Diffusion Posterior Sampling Reconstruction in Low-Dose CT.

Area of Science:

• Medical Imaging
• Computational Imaging
• Radiology

Background:

• Low-dose computed tomography (CT) aims to reduce radiation exposure while maintaining image quality.
• Algorithmic solutions, including deep learning generative priors like Diffusion Posterior Sampling (DPS), are crucial for noise control in low-dose CT.
• A key question is how to best achieve low dose: by reducing the number of projection views or the photon fluence per view.

Purpose of the Study:

• To systematically investigate the impact of acquisition parameters (number of views and incident photons per view) on Diffusion Posterior Sampling (DPS)-based CT reconstruction.
• To determine the optimal strategy for achieving low-dose CT protocols by comparing sparse sampling versus reduced fluence per projection.
• To provide a framework for optimizing CT acquisition protocols for reduced radiation doses.

Main Methods:

• A simulation study was conducted to evaluate DPS-based CT reconstruction.
• A 2D parameter sweep across various combinations of the number of views (n_view) and incident photons per view (I_0) was performed.
• Reconstruction quality was assessed using Peak Signal-to-Noise Ratio (PSNR), bias, and posterior sample variability, comparing results for equivalent total incident photons (TIP).

Main Results:

• The number of views significantly influences CT image quality, with a critical performance drop when the number of views is too low.
• Joint optimization of photon fluence per view and the number of views yields superior results compared to optimizing either parameter alone.
• An increasing number of views is generally required as the total incident photon fluence increases to maintain optimal performance.

Conclusions:

• The number of views is a critical parameter for low-dose CT reconstruction quality using DPS.
• Optimizing CT acquisition protocols requires a joint strategy that adapts both photon fluence per view and the number of views (sparsity).
• These findings offer a data-driven approach to designing low-dose CT protocols that balance radiation dose and image fidelity.