[Brainmap]: Enhao Gong - Improve Deep Learning based MRI reconstruction with Generative Adversarial Network(GAN)

Abstract:

Magnetic Resonance Imaging (MRI) reconstruction is a severely ill-posed inversion task requiring intensive computations. Recently various methods have been proposed to apply different Deep Learning models for more efficient and accurate MRI reconstruction. The first part of the talk will briefly introduce our works in this area, specifically for the application of multi-contrast MRI reconstruction and enhanced reconstruction of ASL. 

Although Deep Learning based MRI reconstruction attracts great attention and enthusiasm, there are still open questions on how to choose network models, design network structure and optimize cost function in learning. The second part of the talk will focus on our recent work on a new design of Deep Learning based MRI reconstruction framework with generative adversarial network (GAN). In the new framework, one network is trained for reconstruction by learning manifold projection and aliasing removal, while the other network is jointly trained to discriminate the reconstruction quality. 
Evaluated on a large contrast-enhanced MR dataset with both quantitative metrics and radiologists' ratings, the proposed method, GANCS, demonstrates the superior reconstruction performance compared with both much slower Compressed Sensing (CS) reconstruction and Deep Learning reconstruction model trained with pixel-wise MSE loss.
 
About the Speaker:
Enhao Gong is a PhD candidate in Electrical Engineering at Stanford. His research focus is on applying machine learning, deep learning and optimization for medical imaging reconstruction and processing. Specifically, he is working on fast Magnetic Resonance Imaging (MRI) algorithms, multi-contrast neuroimaging applications (MRI, PET/MR). He is working with Professor John Pauly in EE and Professor Greg Zaharchuk in Radiology. 
Recently he is working to bridge deep learning methods with under-sampled MRI reconstruction, such as enhancing ASL with Deep Learning and multi-contrast information, solving water-fat separation using Deep Learning framework as well as using Deep Generative Adversarial Network (GAN) for Compressed Sensing MRI.
He has also founded Polarr, a start-up working on computer vision and applying deep learning to mobile applications. He has worked at Qualcomm for optical recognition technologies and Philips Healthcare for MRI reconstruction. Before joining the MS/PhD program at Stanford, Enhao Gong graduated from Biomedical Engineering at Tsinghua University in China where he was working on optical imaging, neuroimaging and Brain Computer Interface.