The generation of realistically simulated photoacoustic (PA) images with ground truth labels for optical and acoustic properties has become a critical method for training and validating neural networks for PA imaging. As state-of-the-art model-based simulations often suffer from various inaccuracies, unsupervised domain transfer methods have been recently proposed to enhance the quality of model-based simulations. The validation of these methods, however, is challenging as there are no reliable labels for absorption or oxygen saturation in vivo. In this work, we examine various domain shifts between simulations and real images such as simulating the wrong noise model, inaccuracies in modeling the digital device twin or erroneous assumptions on tissue composition. We show in silico how a Cycle GAN, unsupervised image-to-image translation networks (UNIT) and a conditional invertible neural network handle these domain shifts and what their consequences are for blood oxygen saturation estimation.
|