Physics-informed neural networks for quantitative photoacoustic imaging

Quantitative photoacoustic imaging remains an open challenge, particularly due to the limited translatability of methods that achieve excellent results in silico to real-world settings. We present a physics-informed neural network approach that integrates a differentiable discrete backprojection into the model architecture. This combines the robustness of model-based inversions with the accuracy of data-driven model corrections. We train the network using digital phantoms and anatomically realistic mouse simulations that are calibrated with a set of real and numerical phantom data. Our results indicate improved generalisability of the quantification network when using a physics-informed network architecture, and we validate the model performance on a diverse set of in silico and experimental phantom and in vivo mouse data.