A comparative study of diffusion-based reconstruction frameworks for photoacoustic tomography.

Area of Science:

• Medical Imaging
• Computational Imaging
• Image Reconstruction

Background:

• Photoacoustic tomography (PAT) reconstruction is an ill-posed inverse problem.
• Limited-view acquisition in PAT introduces artifacts and degrades structural information.
• Diffusion-based generative models offer a promising approach for incorporating learned priors into image reconstruction.

Purpose of the Study:

• To systematically compare three diffusion-based frameworks (DDPM, DDIM, score-based models) for PAT reconstruction.
• To evaluate performance under both full-view and limited-view conditions.
• To analyze the impact of physics-informed measurement-consistency refinement on reconstruction fidelity.

Main Methods:

• Comparison of denoising diffusion probabilistic models (DDPM), denoising diffusion implicit models (DDIM), and score-based models.
• Testing on synthetic and anatomical data under full-view and limited-view acquisition.
• Analysis of physics-informed measurement-consistency refinement during sampling.

Main Results:

• DDPM yields accurate full-view reconstructions but degrades in limited-view settings without refinement.
• DDIM provides comparable quality with fewer sampling steps.
• Score-based models show superior performance across all settings, enhancing structural fidelity and perceptual quality, despite more sampling steps.

Conclusions:

• Score-based diffusion models are effective for challenging PAT reconstruction, especially under limited views.
• Physics-informed refinement is crucial for DDPM performance in limited-view scenarios.
• The choice of diffusion model impacts reconstruction quality and computational cost in PAT.