Abstract
Photoacoustic tomography (PAT) reconstruction is a highly ill-posed inverse problem, particularly under limited-view acquisition, that leads to severe artifacts and loss of structural information. Diffusion-based generative models have recently been explored as a means of incorporating learned priors into image reconstruction. In this study, we systematically compare three representative diffusion-based frameworks: denoising diffusion probabilistic models (DDPM), denoising diffusion implicit models (DDIM), and score-based models for PAT reconstruction under full-view and limited-view conditions using both synthetic and anatomical data. Further, the role of physics-informed measurement-consistency refinement applied during sampling and its impact on reconstruction fidelity across varying degrees of data is analyzed. The results indicate that while DDPM provides accurate reconstructions under full-view acquisition, its performance degrades under limited-view conditions without an explicit refinement step. Similarly, DDIM achieves comparable reconstruction quality while requiring fewer sampling steps. In contrast, score-based models demonstrate consistently better performance across acquisition settings, yielding improved structural fidelity and perceptual quality, at the expense of a higher number of sampling timesteps. These findings provide practical insight into different diffusion formulations and offer guidance for deploying diffusion-based reconstruction methods in challenging PAT imaging scenarios.