Preview

PDF, English Download (8MB) | Terms of use

Abstract

Photoacoustic (PA) imaging is a promising and emerging technique for detection and characterization of diseases, such as superficial tumors. It combines the strengths of optical imaging and ultrasound: high optical contrast as well as high imaging depth and spatial resolution. Handheld PA devices, which are popular due to their versatility and ease of use, only provide a limited field of view caused by their sensor geometry. This restrain in hardware as well as the optical and acoustical attenuation in tissue limit the ability to reconstruct the initial pressure distribution, which is necessary for quantification of the underlying tissue. Today only approximations can be reconstructed with state-of-the-art methods. In this master thesis a novel approach for reconstructing the initial pressure distribution in tissue from limited view PA data is presented. It is based on machine learning, using a fully convolutional deep neural network with a U-Net-like architecture. The data necessary to train this network and validate the approach was generated in silico. The Initial pressure distribution of single and multiple blood vessels and the propagation of ultrasound was simulated in tissue. The results show a promising path to solve the acoustic inverse problem and offer a quantitative and qualitative improvement over state-of-the-art techniques. The method developed in this thesis can be a further step on the road towards clinical quantitative photoacoustic imaging.