Speaker
Description
The acoustic inverse problem in photoacoustic (PA) imaging is ill-posed under typical measurement configurations. We propose a hard-constrained physics-embedded neural network that directly embeds a universal backprojection module without learnable weights, surrounded by a learnable FFT filter and a learnable U-Net projection-combination module. This design enables quantitatively accurate reconstruction of initial pressure distributions even with limited training data.We benchmarked various learning-based PA reconstruction schemes, evaluating data efficiency and generalisability by training on 10, 50, 100, 200, and 300 simulated phantom datasets for 25 epochs with matched parameter counts. Quantitative accuracy of the reconstructed initial pressures was validated on simulated data with increasing difficulty and further tested on experimental data both from phantoms and mouse measurements.Using a mix of soft and embedded physics constraints consistently outperforms fully black-box models, already achieving high quantitative accuracy with as few as 50 training examples and showing robust performance on increasingly out-of-distribution situations. Models lacking physics priors during training failed to generalise within the training budget. Embedding a non-trainable physics module improves data efficiency and quantitative reliability, making the approach viable for targeted photoacoustic applications or uncommon detection geometries, where simulating large training datasets is computationally prohibitive.