| Abstract: |
| Optoacoustic mesoscopy is a biomedical imaging technique that provides images of the microvasculature of the skin. The skin tissue is illuminated with a short laser light pulse. Due to the optoacoustic effect this results in the generation of pressure waves, which leads to the inverse problem of reconstructing an image, representing the absorption properties of the tissue, from ultrasound measurements. We focus on a spherically focused ultrasound transducer performing a raster-scan on the skin surface and use a model-based image reconstruction approach. It consists of computing the forward model matrix describing the measurement process and using variational regularization to solve an optimization problem incorporating the forward model. A main challenge of the model-based reconstruction approach for 3D images is the excessive memory demand and the computation time. We reduce both by exploiting the symmetry of the ultrasound transducer and using a stochastic proximal gradient algorithm for iterative image reconstruction. In each iteration we randomly choose a certain number of model matrix rows and corresponding signal values for which we perform a gradient descent step on the data fidelity term and a proximal step on the regularization term. Furthermore, we use Langevin sampling to enable Bayesian uncertainty quantification. |
|