Deep learning filtered back projection for photoacoustic tomography

Image reconstruction in photoacoustic tomography (PAT) faces significant challenges due to incomplete projection data. To address this problem, we propose a physics-driven, deep learning-based filtered back-projection algorithm (dFBP). The dFBP framework is systematically designed based on the principles of the traditional analytical filtered back-projection algorithm and consists of a filtering module, a back-projection module, and a fusion module arranged in a cascaded architecture. Numerical simulations show that dFBP-based PAT accurately reconstructs high-quality, artifact-suppressed images from sparse, limited, and acoustically heterogeneous projection data.