High-fidelity dual-domain photoacoustic data synthesis with latent diffusion models

Photoacoustic tomography (PAT) is widely utilized in biomedical imaging due to its high optical contrast and deep acoustic penetration. However, the scarcity of high-quality datasets hinders the application of deep learning in this field. To address this, we propose a latent diffusion model (LDM)-based dual-domain data generation framework, which consists of two independently-trained synthesis pathways, that synthesizes photoacoustic data in both image and projection domains. Our method maps raw data into a downsampled latent space, preserving essential structural and semantic information while significantly reducing modeling dimensionality. The two pathways are trained separately to model their corresponding data distributions. To validate the framework, the generated data was used to train a reconstruction network for sparse-view, limited-view, and bandwidth-limited tasks. Experimental results show that models trained with synthetic data achieve performance comparable to reference data training in terms of structural fidelity, artifact suppression, and overall image quality. Quantitative evaluation via PSNR and SSIM metrics further confirms the high usability of the generated data. This study represents the first systematic exploration of dual-domain generation in PAT, offering a robust strategy for deep learning research under medical data constraints.