Iterative Diffusion-Refined Neural Attenuation Fields for Multi-Source Stationary CT Reconstruction: NAF Meets Diffusion Model

TL;DR

The paper tackles ultra-sparse-view reconstruction in multi-source stationary CT by coupling Neural Attenuation Fields (NAF) with diffusion priors in the DR-domain. It introduces Diff-NAF, an iterative framework that alternates NAF training with diffusion-based refinement of simulated projections, guided by an Adaptive Projection Synthesis strategy (APGPS) and a Diffusion-driven Reuse Projection Refinement (DRPR) module featuring Dynamic Range Adaptive Transformation (DRAT). Through iterative pseudo-labeling, refined projections progressively improve data fidelity and structural coherence, yielding superior PSNR/SSIM on simulated and real datasets compared to state-of-the-art methods. The work demonstrates the benefit of integrating physics-based forward modeling with learned priors for fast, robust 3D CT reconstruction under severe undersampling, with potential impact on clinical and industrial rapid-imaging workflows.

Abstract

Multi-source stationary computed tomography (CT) has recently attracted attention for its ability to achieve rapid image reconstruction, making it suitable for time-sensitive clinical and industrial applications. However, practical systems are often constrained by ultra-sparse-view sampling, which significantly degrades reconstruction quality. Traditional methods struggle under ultra-sparse-view settings, where interpolation becomes inaccurate and the resulting reconstructions are unsatisfactory. To address this challenge, this study proposes Diffusion-Refined Neural Attenuation Fields (Diff-NAF), an iterative framework tailored for multi-source stationary CT under ultra-sparse-view conditions. Diff-NAF combines a Neural Attenuation Field representation with a dual-branch conditional diffusion model. The process begins by training an initial NAF using ultra-sparse-view projections. New projections are then generated through an Angle-Prior Guided Projection Synthesis strategy that exploits inter view priors, and are subsequently refined by a Diffusion-driven Reuse Projection Refinement Module. The refined projections are incorporated as pseudo-labels into the training set for the next iteration. Through iterative refinement, Diff-NAF progressively enhances projection completeness and reconstruction fidelity under ultra-sparse-view conditions, ultimately yielding high-quality CT reconstructions. Experimental results on multiple simulated 3D CT volumes and real projection data demonstrate that Diff-NAF achieves the best performance under ultra-sparse-view conditions.

Figures (12)

• Figure 1: Overview of the reconstruction tasks and our method. (a) The process of performing 3D reconstruction from DR projections using NAF. (b) The process of reconstructing image slices from sinograms using the diffusion model. (c) Our proposed framework: DR projections are used as input, followed by an iterative process between NAF and the diffusion method, ultimately producing a high-quality 3D reconstruction.
• Figure 2: Schematic of a multi-source stationary CT configuration. Multiple fixed X-ray sources ($S_1$–$S_n$) and corresponding detectors ($D_1$–$D_n$) are arranged around the object, where $n$ is typically less than 30. During acquisition, the sources are activated sequentially or in a time-multiplexed sequence, and the detectors record DR projections from different directions.
• Figure 3: Motivation schematic. We consider a setting where NAF is trained on full 3D volumes, whereas the diffusion model is trained on axial slices. (a) On axial slices, the diffusion model delivers higher quality and greater stability than NAF. (b) On the coronal plane, NAF still performs well; however, the diffusion approach—constructed by stacking axial predictions and re-slicing along the coronal axis—exhibits severe inter-slice inconsistencies.
• Figure 4: Overview of the proposed Diff-NAF framework. (a) Iterative reconstruction framework: starting from initial sparse-view projections, the NAF is trained and then enters an iterative process that includes novel-view synthesis, DR projection correction, pseudo-label merging, and NAF retraining, followed by CT reconstruction. (b) NAF: illustrates the training procedure and the synthesis of novel-view projections. (c) shows the training process of the dual-branch conditional diffusion model, which is the main component of the DRPR module.
• Figure 5: Illustration of the dual-branch diffusion correction process. The vertical axis ($\bar{\beta}_t$) denotes the denoising weight, and the horizontal axis ($\bar{\alpha}_t$) denotes the residual removal weight. The curve shows the evolution of DR correction, demonstrating the simultaneous removal of residuals and noise during the diffusion process.