Learning Tissue Geometries for Photoacoustic Image Analysis

Photoacoustic imaging (PAI) holds great promise as a non-ionizing imaging modality, allowing insight into physiological tissue properties. The estimation of physiological tissue properties with PAI, however, involves two challenges: (1) there is a lack of ground truth labels for these properties in vivo, and (2) the estimation requires the solution of a non-linear, ill-posed inverse problem. While deep learning (DL) approaches trained on simulations are promising to address the inverse problem, their applicability is hampered by the current sim-to-real domain gap in PAI. This thesis aimed to explore novel data-driven methods to enhance the realism of PAI simulations (learning-to-simulate). The specific research focus was placed on methods generating tissue geometries covering a variety of different tissue types and morphologies and representing a key component in PAI simulation approaches.