Beyond the Patterns - Mert Sabuncu (Cornell U): Deep Learning for Compressed Imaging

We have the great honor to welcome Mert Sabuncu to our lab for an invited presentation! Abstract: Imaging techniques such as MRI can be accelerated by sampling below the Shannon-Nyquist rate via compressed sensing. In this talk, I will consider the use of deep learning methods for this problem. First, I will present our approach for Learning-based Optimization of the Under-sampling PattErn, or LOUPE. For a given sparsity constraint, LOUPE optimizes the under-sampling pattern and reconstruction model simultaneously, using a computationally efficient end-to-end deep learning strategy. Our experiments with MRI and microscopy demonstrate that LOUPE-derived patterns yield significantly more accurate reconstructions compared to standard under-sampling schemes. I will then switch gears and focus on the reconstruction problem only and presents some deep-learning based innovations that we have recently proposed, including the use of hyper-networks that give end-users to ability to choose from multiple reconstructions