You are here

Share:

Progressive and Multipath Neural Network for Medical Image Segmentation

Summary
Researchers at the National Institutes of Health Clinical Center (NIHCC) developed a technology that improves segmentation detail levels for anatomical structures in medical images through a new, deep learning approach. Difficult anatomical features, often segmented incorrectly with other image segmentation methods, are correctly segmented and identified using this novel technology.
NIH Reference Number
E-109-2017
Product Type
Keywords
  • Segmentation, Holistically-nested networks (HNNs), Medical Imaging, Automated Medical Image, Medical Imaging Processing, Image Segmentation, Deep Learning, Computer Assisted Diagnostic, Diagnostics, Coarsening Resolution
Collaboration Opportunity
This invention is available for licensing and co-development.
Contact
Description of Technology

The technology relates to software, systems, and methods for automated medical image segmentation via deep learning. Standard holistically-nested networks (HNNs) used in computer vision and medical imaging for segmentation and edge detection have a problem with coarsening resolution. The described technology addresses this.
 
Simple, progressive multipath connections enhance the standard HNN model.  Processing in this new model achieves improved segmentations detail levels without need for additional model parameters.  Variations in appearance are handled without being affected by variations in shape for a given image feature. Early experiments improved segmentation masks compared to standard HNN. Anatomical regions, often misidentified, are correctly captured via this approach.

This technology is generalizable to image segmentation across multiple imaging modalities, including CT, MRI and X-ray.  The improved image segmentation detail levels may have clinically relevant applications to high variability image features, such as tumors and pathologies of the lung. This technology has potential to improve diagnostic capabilities and treatment outcomes in many disease conditions. 

Potential Commercial Applications
  • Computer assisted diagnostics and medical imaging processing
  • Tumor imaging/diagnosis 
  • Imaging pathologies of the lung
  • Diagnostic capabilities for a range of disease conditions; e.g., cardiovascular and neurodegenerative
  • Providing framework to develop more cost-effective health care
Competitive Advantages
  • Improved detail in automated medical image segmentation without need for additional training or processing of images on other parameters. 
  • Generalizable to image segmentation across multiple imaging modalities, including CT, MRI and X-ray
  • Improved segmentations detail levels without need for additional model parameters
  • Improved segmentation masks compared to standard HNN
Inventor(s)

Le Lu (NIHCC), Daniel J Mollura (NIHCC), Ronald M Summers (NIHCC), Adam P Harrison (NIHCC), Ziyue Xu

Development Stage
Patent Status
  • U.S. Provisional: U.S. Provisional Patent Application Number 62/516,948, Filed 08 Jun 2017
Posted
Monday, October 23, 2017