Solving the visibility problem in photoacoustic imaging with a deep learning approach providing prediction uncertainties

Photoacoustic imaging is a promising biomedical imaging modality providing optical contrasts at depth. Conventional reconstructions suffer from the limited view and bandwidth of the ultrasound transducers. As a result, structures elongated in the axis of the probe or large compared to the point spread function of the system are not fully recovered. A deep learning approach is proposed to handle these problems and is demonstrated both in simulations and in experiments on a multi-scale model of leaf skeleton. We employed an experimental approach to build the training and the test sets using photographs of the samples as ground truth images. Reconstructions produced by the neural network show a greatly improved image quality as compared to conventional approaches. In addition, this work aimed at quantifying the reliability of the neural network predictions. To achieve this, the dropout Monte-Carlo procedure is applied to estimate a pixel-wise degree of confidence on each predicted picture. This degree of confidence is well correlated to the variability observed in certain regions, among a stack of subsequently acquired frames. Last, we address the possibility to use transfer learning with simulated data in order to drastically limit the size of the experimental dataset.