Gordon Lecture: Deep Learning MR Reconstruction from Missing Data

Jong Chul Ye is currently KAIST Endowed Chair Professor and Professor of the Dept. of Bio/Brain Engineering and Adjunct Professor at Dept. of Mathematical Sciences of Korea Advanced Institute of Science and Technology (KAIST), Korea. He received the B.Sc. and M.Sc. degrees from Seoul National University, Korea, and the Ph.D. from Purdue University, West Lafayette.
Below is a summary of his presentation

Recently, deep learning approaches with various network architectures have achieved significant performance improvement over existing iterative reconstruction methods in accelerated MRI problems.  However, it is still unclear why these deep learning architectures work for specific problems. Moreover, in contrast to the usual evolution of signal processing theory around the classical theories, the link between deep learning and the classical image processing approaches are not yet well understood. In this talk, Dr. Ye reviewed the recent advances of deep learning approaches for accelerated MRI and their link between compressed sensing approaches. 

In particular, Dr. Ye first reviewed the variational neural network that was first proposed in MR field,  and the popular feed-forward neural network approaches using U-Net, which can remove undersampling artifacts from the aliasing artifact corrupted image. Then, he reviewed several advanced approaches such as AUTOMAP, CascadeNet, KiKi-Net, MoDL, etc. Finally, he demonstrated that the neural network approaches can be directly implemented in k-space domain to interpolate the missing k-space data.   In order to explore the theoretical origin of the success of the neural network for accelerated MRI, Dr. Ye reviewed some of the mathematical principles that have been proposed to explain the neural networks for inverse problems, which includes unfolding, convolution framelets, etc.  Then, he introduced recent mathematical discovery of the expressivity, generalization power and optimization landscape that give us hint to understand the power of AI for accelerated MRI.