Deep learning-enhanced 3D real-time photoacoustic imaging using experimental ground truths obtained from fluctuation imaging.

Abstract

3D conventional photoacoustic (PA) imaging often suffers from visibility artifacts caused by the limited bandwidth and constrained viewing angles of ultrasound transducers, as well as the use of sparse arrays. PA fluctuation imaging (PAFI), which leverages signal variations due to blood flow, compensates for these visibility artifacts at the cost of temporal resolution. Deep learning (DL)--based photoacoustic image enhancement has previously demonstrated strong potential for improved reconstruction at a high temporal resolution. However, generating an experimental training dataset remains problematic. Herein, we propose creating an experimental training dataset based on single-shot 3D PA images (input) and corresponding PAFI images (ground truth) of chicken embryo vasculature, which is used to train a 3D ResU-Net neural network. The trained DL-PAFI network predictions on new experimental test images reveal effective improvement in visibility and contrast. We observe, however, that the output image resolution is lower than that of PAFI. Importantly, incorporating only experimental data into training already yields a good performance, while pre-training with simulated examples improves the overall accuracy. Additionally, we demonstrate the feasibility of real-time rendering and present preliminary in vivo predictions in mice, generated by the network trained exclusively on chicken embryo vasculature. These findings suggest the potential for achieving real-time, artifact-free 3D PA imaging with sparse arrays, adaptable to various in vivo applications.