Technology ID

General-purpose Deep Learning Image Denoising Based on Magnetic Resonance Imaging Physics

Lead Inventor
Xue, Hui (NHLBI)
Kellman, Peter (NHLBI)
Software / Apps
Development Stages
Pre-clinical (in vivo)
Research Products
Computational models/software
Lead IC

This technology includes a novel method to train deep learning convolution neural network model to improve the signal-noise-ratio for the magnetic resonance (MR) imaging. The novelty lies on the fact that actual MR imaging physics information is used in the deep learning training. The resulting model achieves significant signal-to-noise ratio (SNR) improved for different acceleration factors in MR imaging. The resulting model can be used for many body anatomies (e.g., brain, heart, liver, spine, etc.) to significantly improve the SNR. This solution is fast enough to be used clinically and has already been implemented on MR scanners.

Commercial Applications
  • The resulting model is general-purposed and is applicable to many MR imaging applications and different human anatomy.
  • This is a technique which can be used in every clinical MR scan to improve SNR.
Competitive Advantages
  • Better/correct way to do deep learning magnetic resonance denoising
  • Ability to break g-factor barrier for up to 5x acceleration
  • Ability to work across scanners/anatomy/imaging protocols
  • No assumption on deep learning model architecture
Licensing Contact:
Kolesnitchenko, Vincent