Physics-informed neural network for photoacoustic tomography reconstruction.

Conventional photoacoustic tomography (PAT) reconstruction algorithms often suffer from performance degradation under complex media and non-ideal detection conditions. To tackle these limitations, we propose a novel reconstruction method based on physics-informed neural networks (PINNs). Approach: The network model is optimized with constraints from observational data, wave equation, and initial conditions, to reconstruct the spatial distribution of initial acoustic pressure. The PINN framework further incorporates time reversal method to improve computational efficiency. Main results: Comprehensive PAT simulations demonstrate that the PINN achieves superior reconstruction quality compared with conventional algorithms. The PINN exhibits favorable adaptivity to both heterogeneous and lossy media, and strong robustness against noisy measurements, sparse sensor layouts and limited-view detection geometries. Significance: The PINN integrates the physical interpretability and the powerful capability of neural networks to represent continuous acoustic pressure fields. It eliminates the requirement for training datasets, offering broad application potential for PAT.