Technology ID

A Machine Learning Strategy to Improve the Fidelity of Imaging Time-Varying Signals to Improve Clinical Imaging

Lead Inventor
Hsu, Li-Yueh (NHLBI)
Bui, Vy (NHLBI)
Jacobs, Matthew (NHLBI)
Benovoy, Mitchel (Corstem Inc)
Arai, Andrew (NHLBI)
Software / Apps
Research Materials
Research Products
Computational models/software
Lead IC
This technology includes a new technique to improve the fidelity of time-varying signals acquired in the dynamic contrast enhanced (DCE) imaging. This technique enhances the time-varying signals in a given DCE image series through deep convolutional neural networks (CNN) to learn the relationship of signal versus contrast concentration from other series of different contrast doses. The method includes an automatic framework to detect target regions of interest (ROI), match key signal-time frames, and register anatomical structures in different DCE series to train the CNN and produce an image series with improved signal fidelity. Depending on specific ROI in the DCE images, signal fidelity may include the linearity of the signal versus contrast concentration, or measurements of image quality. The method automatically selects key signature frames in two calibrated DCE image series and matches only those frames in the training. Once properly trained, the inference network can be applied to the entire DCE image series to generate a synthetic DCE image series that has improved fidelity.
Commercial Applications
Application to clinical DCE imaging application to improve quantitative analysis of time-varying signals and qualitative interpretation of the DCE image, as well as reduce contrast dose required in DCE imaging for clinical diagnosis.

Competitive Advantages
The propose technique not only can improve region specific time-varying signals in a DCE image series, it can also reduce the amount of required matched image data to significantly reduce the CNN training time.
Licensing Contact:
Shmilovich, Michael