Conventional photoacoustic (PA) imaging suffers from visibility artefacts due to limitations in ultrasound transducer bandwidth, viewing angles, and the use of sparse arrays. PA fluctuation imaging (PAFI), exploiting the signal changes due to blood flow, compensates for these artefacts, at the cost of temporal resolution.
Our study addresses this limitation employing a deep learning approach in which PAFI images serve as ground truths for training a 3D neural network to obtain real-time single-shot artefact-free images.
Following a pre-training with simulated examples, a 3D-ResUnet network was trained with 90 PA chicken embryo vasculature volumes as input and corresponding PAFI as ground truths. Notably, inclusion of experimental data significantly improves predictions over simulation-only training, even accounting for transducer angular filtering.
Furthermore, applying the same network exclusively trained in-ovo to predict the femoral artery in mice demonstrates the potential of this method for real-time, full-visibility multispectral PA imaging in vivo using sparse arrays.
|