Model-based Fully Dense UNet for Image Enhancement in Software-defined Optoacoustic Tomography

A deep neural network architecture for improving the performance of a software-defined optoacoustic tomography device is presented. Our approach is a hybrid one, in the sense that a powerful data-driven architecture (a FD-UNet) is combined with a structure that exploits model-guided information, in the form of the forward and adjoint operators of the acoustic problem. Besides that, the findings of a previous work on the noise and other effects on the measured sinograms are also exploited, in order to make the structure more robust in the task of correcting the artifacts that are typically introduced in the reconstructed images. The proposed solution is numerically trained and evaluated. In terms of the average mean square error over the testing data-set, our approach shows better performance than well-established reconstruction algorithms in the field of optoacustic tomography. A series of examples shows that this superior performance also holds with respect to other reconstruction image quality measures.