Seminar

Medical Physics Seminar – Monday, April 8, 2019

Low-Cost Fetal Magnetocardiography

speaker

Sarah Strand (student of Dr. Ronald Wakai)

Fetal magnetocardiography (fMCG) is a highly effective technique for the evaluation of fetuses with life-threatening arrhythmia; however, its translation to the clinic has been hindered by the high cost of superconducting quantum interference device (SQUID) magnetometers and magnetically-shielded rooms (MSRs). A SQUID-based system, including MSR, costs approximately $1M. This situation is about to change due to the recent availability of high-performance optically-pumped magnetometers (OPMs). The required components of a OPM-based system (8 sensors and person-sized shield) can be purchased for as little as $100k. In addition to reduced sensor cost, their small size permits the use of person-sized magnetic shields. This study compares the performance of an OPM fMCG system operating in a person-sized, cylindrical magnetic shield (CMS) with a SQUID fMCG system operating in an MSR. The purpose of this work is to demonstrate that an OPM-based fMCG system can per-form as well as or better than a SQUID-based system at a small fraction of the cost. OPM technology has the potential to make fMCG much more affordable and widely available.


Deep Learning Uncertainty Machnism and Bayesian Deep Learning Models for MR Neuroimaging

speaker

Gengyan Zhao (student of Beth Meyerand)

In the recent past, tremendous progress has been made in deep learning as a result of the revival of deep neural net-works as well as the rapid advance of parallel computing techniques. The proposal of Bayesian deep learning further enabled the ability of uncertainty generation in deep learning prediction. Among deep learning models, convolutional neural network (CNN) has proven to be useful in a broad range of computer vision applications, outperforming traditional state-of-the-art methods. Stacked autoencoders (SAE) are good at dealing with non-image data with high-dimensional feature space. Generative adversarial network (GAN) is a powerful tool in various image synthesis tasks, including image to image translation and image reconstruction. This presentation will outline the development of deep learning and Bayesian deep learning models for image segmentation, classification, feature ex-traction, image synthesis and related uncertainty generation in MR neuroimaging.


Location: 1325 HSLC (Health Sciences Learning Center), 750 Highland Ave., Madison, WI 53705 - USA

Time: 4:00pm-5:00pm