Image reconstruction in tomography with incomplete data using microlocal aware deep learning

Abstract:

The talk outlines recent progress in developing domain adapted deep neural networks for reconstruction in tomography with incomplete data (more precisely limited angle tomography). The key idea is to develop a deep neural network architecture that is aware of which edges in the image that are visible and how they relate to edges in noisy incomplete tomographic data. A key component is the microlocal correspondence, which is a mathematical description of how a deep neural network for reconstruction transforms such visible edges. Another is to design a robust deep learning based extraction of edges in digitised image and data. These components are combined to handcraft a deep neural network architecture for image reconstruction that is specifically adapted for limited angle tomography.