3D Reconstruction in photoacoustic imaging assisted by deep-learning

Photoacoustic imaging (PA) is an emerging biomedical modality consisting in the emission of laser light, which, when is absorbed by the tissue components, generates ultrasound waves. After reception, the signals PA are used to provide an image by reconstruction algorithm, as the delay-and-sum beamforming. In the vast field of biomedical imaging, this modality is really promising, as it permits to image the tissue optical properties at depths with interesting resolutions and can provide images of the optical absorption with specific molecular contrast which can be enhanced by spectroscopy. In particular, the omnipresence of haemoglobin in living tissues allows the imaging of the microvasculature, which is one of the most important uses of photoacoustic imaging. However, conventional PA imaging systems are limited by low contrast and visibility artefacts that arise from coherence of PA waves and characteristics of the detection system such as geometry and frequency bandwidth. Limited bandwidth artefacts occur when the central part of the reconstructed object are not visible because ultrasound detectors filter out low-frequency components of PA waves emitted by large absorbers (large as compared to the detection wavelength range). Limited-view artefacts occur when the coherent acoustic waves being directional cannot be measured if they don’t reach the probe. A first dynamic technique to solve these visibility problems is the photoacoustic fluctuation imaging that exploits the natural fluctuation of the blood flow to reconstruct the total visibility of the vessels. It was shown on a 2D imaging system that the reconstruction quality could be also enhanced thanks to a deep learning algorithm trained on simulated and experimental data. The main objective of this thesis is to transfer these results on a 3D environment by elaborating a neural network able to process 3D volumes. In the first part of this project we focus on programming an user-friendly and reliable acquisition system to collect time signals from a chicken embryo model. Its chorioallantoic membrane is, in fact, an optimal in vivo model to study blood vessels thanks to its visibility, accessibility and rapid developmental growth. From radio-frequency signals, we aim at developing a procedure to create an experimental training-set consisting of conventional PA images as input and PA fluctuation images, free of visibility problems, used as ground truth. Our deep learning approach includes the pre-training the network with simulated data and a following training with the experimental data to find a model that permits, from inputs never trained and not known by the network, to predict an output free of artefacts. The correctness and the robustness of this prediction is finally verified by a correlation measure.