Robust Photoacoustic Beamforming Using Dense Convolutional Neural Networks.

Photoacoustic (PA) is a promising technology for imaging of endogenous tissue chromophores and exogenous contrast agents in a wide range of clinical applications. The imaging technique is based on excitation of a tissue sample using short light pulse, followed by acquisition of the resultant acoustic signal using an ultrasound (US) transducer. To reconstruct an image of the tissue from the received US signals, the most common approach is to use the delay-and-sum (DAS) beamforming technique that assumes a wave propagation with a constant speed of sound. Unfortunately, such assumption often leads to artifacts such as sidelobes and tissue aberration; in addition, the image resolution is degraded. With an aim to improve the PA image reconstruction, in this work, we propose a deep convolutional neural networks-based beamforming approach that uses a set of densely connected convolutional layers with dilated convolution at higher layers. To train the network, we use simulated images with various sizes and contrasts of target objects, and subsequently simulating the PA effect to obtain the raw US signals at an US transducer. We test the network on an independent set of 1,500 simulated images and we achieve a mean peak-to-signal-ratio of 38.7 dB between the estimated and reference images. In addition, a comparison of our approach with the DAS beamforming technique indicates a statistical significant improvement of the proposed technique.