Presented By O’Reilly and Intel AI
Put AI to Work
April 29-30, 2018: Training
April 30-May 2, 2018: Tutorials & Conference
New York, NY

Deep learning and AI is making clinical neuroimaging faster, safer, and smarter

Enhao Gong (Stanford University | Subtle Medical), Greg Zaharchuk (Stanford University)
2:35pm–3:15pm Tuesday, May 1, 2018

Who is this presentation for?

  • Executives in healthcare

What you'll learn

  • Learn how AI and deep learning has already introduce disruptive changes to healthcare services and clinical workflow
  • Discover how AI and deep learning can not only assist diagnosis and treatment planning based on medical images but can also enable faster, safer, cheaper, and smarter medical imaging services

Description

What is the impact of AI and deep learning on clinical workflows? Enhao Gong and Greg Zaharchuk offer an overview of AI and deep learning technologies invented at Stanford and applied in the clinical neuroimaging workflow at Stanford Hospital, where they have provided faster, safer, cheaper, and smarter medical imaging and treatment decision making.

Specific use cases include:

  • A multiscale 3D network that enables lesion outcome prediction for stroke: A lot of methods have been proposed for brain lesion detection and segmentations using machine learning techniques. It is more sophisticated than common computer vision tasks since the estimation of treatment outcomes are not merely determined by lesions captured by current MR images. Researchers at Stanford developed a multiscale 3D neural network for predicting the outcome of stroke patients and final lesion contour shown on day-90 scans by using acute stroke neuro-MRI scans taken on day-0. The algorithm won the MICCAI 2017 ISLES competition. Enhao and Greg also introduce other neuroimaging segmentation and detection projects.
  • Deep learning and GAN that enables >95% reduction in radiation for functional medical imaging: Various medical imaging modalities (mammography, CT, PET, etc.) use the interactions between tissues and radiation to noninvasively visualize anatomy and functions. Positron emission tomography (PET) is a widely used nuclear functional imaging that visualizes the metabolism activities using radioactive tracers for clinical applications, such as diagnosing cancer, heart disease, and neurological disorders. It is a challenging task to reduce radiation and related risks without compromising image quality and diagnosis information. Enhao and Greg share a deep learning method to reconstruct ultra-low-dose PET using residual encoder-decoder network and generative adversarial network. Validated on both clinical oncology datasets and large neuroimaging datasets, the proposed method achieves significantly better results than the state-of-the-art low-dose PET reconstruction methods and preserves diagnostic information while reducing at least 99.5% of the radiation. They demonstrate how using GAN and improved conditional loss functions achieved significantly improved results.
  • Deep learning that enables a 90% reduction in chemical (gadolinium) contrast agent usage in contrast-enhanced MRI: There are increasing concerns globally over the administration of gadolinium-based contrast agents (GBCAs). Enhao and Greg detail a deep learning method that reduces the gadolinium dose in contrast-enhanced MRI (CE-MRI). Trained and evaluated on clinical neuroimaging datasets with reduced contrast usage, both quantitative metrics and radiologists’ ratings showed the proposed method achieved improved synthesis. Synthesized contrast enhancement results using the proposed method have advantages on reduced motion artifacts while preserving the same contrast-enhancement quality compared with current standard results.
Photo of Enhao Gong

Enhao Gong

Stanford University | Subtle Medical

Enhao Gong is a PhD student in electrical engineering at Stanford, where he is advised by John Pauly (electrical engineering) and Greg Zaharchuk (radiology). As founder and researcher at Subtle Medical, he is pushing the performance of deep learning methods to boost the efficiency and value for medical imaging. His research focuses on applying machine learning, deep learning, and optimization for medical imaging reconstruction and processing. Recently, Enhao has been working to bridge deep learning methods with MRI reconstruction, such as enhancing image quality with deep learning and multicontrast information, solving quantitative imaging (water-fat separation, QSM, parameter mapping) using deep learning frameworks, and using generative adversarial networks (GANs) for compressed sensing MRI.

Photo of Greg Zaharchuk

Greg Zaharchuk

Stanford University

Greg Zaharchuk is an associate professor in radiology at Stanford University and a neuroradiologist at Stanford Hospital. His research interests include deep learning applications in neuroimaging, imaging of cerebral hemodynamics with MRI and CT, noninvasive oxygenation measurement with MRI, clinical imaging of cerebrovascular disease, imaging of cervical artery dissection, MR/PET in neuroradiology, and resting-state fMRI for perfusion imaging and stroke.