University PhD Dissertation Defense: Applying Machine Learning/Deep Learning for improved MRI acquisition, reconstruction and quantification, Enhao Gong, Advisor: John Pauly

University PhD Dissertation Defense<br><br>Title: Applying Machine Learning/Deep Learning for improved MRI acquisition, reconstruction and quantification<i><br></i>Speaker: Enhao Gong, Advisor: John Pauly<br>Date: November 6, 2017<br>Time: 1:30pm<br>Location: Glazer Learning Center P083, MSLS Extension, Lucas Center for Imaging (Radiology), <span face="wf_segoe-ui_normal, Segoe UI, Segoe WP, Tahoma, Arial, sans-serif, serif, EmojiFont" style="">1201 Welch Road</span><br><span face="wf_segoe-ui_normal, Segoe UI, Segoe WP, Tahoma, Arial, sans-serif, serif, EmojiFont" style=""><br></span><br><span face="wf_segoe-ui_normal, Segoe UI, Segoe WP, Tahoma, Arial, sans-serif, serif, EmojiFont" style="">Abstract: </span><br>Magnetic Resonance Imaging (MRI) is a powerful medical imaging modality providing fine anatomical details and great tissue contrasts. Variable methods have been proposed to accelerate MRI with under-sampling and regularized reconstruction. However, there are still unsolved problems. Firstly, it is not clear what is the best under-sampling strategy. Secondly, the reconstruction can be a severely ill-posed inversion task due to extreme under-sampling. Commonly used iterative regularized (Compressed Sensing) reconstruction can be time-consuming and not necessary accurate.<br><span face="wf_segoe-ui_normal, Segoe UI, Segoe WP, Tahoma, Arial, sans-serif, serif, EmojiFont" style=""> <p> In the first part, I will talk about the work on optimizing MRI under-sampling by using information from existing scans with similar anatomy. A modified cost function and a stochastic optimization framework is proposed. Together with hardware acceleration, the proposed method can generate a sampling proposal within 5~10sec and enable 20%~40% reduction of resulting reconstruction error.</p> <p> In the second part of the talk, I will discuss on the breakthrough of applying Deep Learning (DL) in MRI reconstruction. Although DL based MRI reconstruction attracts great attention and enthusiasm, there are still open questions on how to design network models and optimize cost function in training. In the proposed DL based MRI reconstruction framework with generative adversarial network (GAN), 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 datasets, the propose method demonstrate superior reconstruction performance and up to 100x speedup, compared with Compressed Sensing (CS) reconstruction and DL reconstruction model trained with pixel-wise loss. </p> The talk will conclude with summary and discussion on research directions to further improve the repeatability and quantification of MRI scans. <br></span>

Date: 
Monday, November 6, 2017 - 1:30pm to 2:30pm
location: 
Glazer Learning Center P083, MSLS Extension, Lucas Center for Imaging (Radiology), 1201 Welch Road