Webinar: Deep Learning-Based Image Reconstruction for MRI
Magnetic resonance imaging (MRI) is a powerful diagnostic tool for visualizing soft tissue anatomy, but physical limits on imaging speed result in MRI exams lasting as long as 90 minutes. The stress of such a lengthy exam has led to attempts to shorten this time, but this can introduce artifacts which obscure relevant anatomy. Attempts have been made to iteratively reconstruct undersampled data into high-quality images using numerical optimization methods, but this is very time-consuming as well.
More recently, it has been shown that deep neural networks can learn to reconstruct undersampled data into images with even higher quality than numerical optimization. Specifically, these deep learning-based (DL) reconstructions are trained to map from the undersampled input to a fully-sampled output in a supervised fashion. However, some works have shown that DL reconstruction methods are unreliable when applied to data unlike the training dataset, and can produce plausible-looking images with textures and anatomical structures which are not real (and are called “hallucinations”).
In this talk, Christopher Michael Sandino will present his work on incorporating physics-based modelling into neural network architectures for DL reconstruction. By leveraging information about the MRI acquisition model throughout the network, undersampled data can be efficiently reconstructed into high-quality images while reducing the chance of hallucinations. These networks are applied to time-resolved MRI, which is commonly used in pediatric patients to determine cardiac health and identify anatomical abnormalities such as congenital heart disease. Christopher will also present his experience with having deployed this technique at Lucile Packard Children’s Hospital, where it is being used everyday to accelerate cardiac MRI exam times by a factor of 12.
About the speaker, Christopher Sandino of Stanford University
Christopher Sandino is a PhD student in the EE Department at Stanford Univ. His research interests span the fields of signal processing, optimization, and AI, with an emphasis on applying his novel techniques to pediatric magnetic resonance imaging. Previously, he majored in EE at USC in Los Angeles. He is a recipient of the National Science Foundation Graduate Research Fellowship, and has held internship positions at the National Institutes of Health, GE Healthcare, and Apple.