Correcting visibility artefacts in photoacoustic imaging with a deep learning approach

Abstract

In conventional photoacoustics (PA) imaging, the finite size and limited-bandwidth of ultrasound transducers often lead to visibility artifacts resulting in a degraded image quality. We propose a reconstruction algorithm based on deep learning to address theses issues. An in vitro vasculature mimicking model has been used in order to show the capability of a conventional neural network to remove these artefacts in an experimental configuration. The deep learning algorithm is trained using couples of PA images and ground truth photographs. The uncertainty of the model prediction is estimated through the Monte Carlo dropout method allowing the display of a pixel-wise degree of confidence. Finally, the interest of using simulation data through transfer learning in order to reduce the size of the experimental dataset is investigated.