Physics-driven deep learning photoacoustic tomography

Photoacoustic tomography (PAT) is a rapidly emerging biomedical imaging modality. To achieve high-performance imaging, signal acquisition in PAT should ideally meet conditions such as full-view detection and dense sampling. However, these circumstances are rarely met in reality, leading to the ill-posed problem of image reconstruction from incomplete projections. Here we propose a physics-driven deep learning-based filtered back projection (dFBP) framework to address this important challenge. The dFBP network is inspired and constructed based on the physical model of the analytical filtered back projection algorithm and consists of a filtering module, a back-projection module, and a fusion module connected in cascade. The dFBP network is driven by physics and is thus interpretable and easy to train while being highly robust and versatile. Numerical and experimental validation on animals and humans show that dFBP-based PAT can achieve direct signal-to-image transformation with enhanced accuracy and reconstruct high-quality, artifact-suppressed images from sparse-view, limited-view, and acoustic heterogeneity-contaminated projections. The proposed dFBP provides a practical solution for high-performance PAT imaging under non-ideal conditions.