Unmet Need
There exists a wide array of strategies for noise reduction in x-ray computed tomography (CT), most of which provide sufficient efficacy in providing a more optimal output. However, current machine learning approaches are developed to provide a singular output of the best solution. This, however, is limiting as certain applications require a degree of control; for example, tuning the tradeoff between denoising and spatial resolution. Thus, there is a need for a control parameter that could represent different possible aspects of image quality in CT images.
Technology Overview
Inventors at Johns Hopkins have developed a novel convolutional neural network (CNN) for applications in low-dose CT that utilizes a hyper-parameter σ for control of balance between noise and bias for specific tasks. The innovation was evaluated using mean squared error and task-based detectability; the results demonstrated that this new parameter allowed for effective control of the level of aggression of noise reduction applied with respect to still maintaining high spatial resolution. While proof of concept work has been completed on tuning of spatial resolution and noise, future development will include tuning of additional parameters (variance, detectability, etc.).
Stage of Development
Schemes have been developed for braiding and proof of concept work has been completed. Inventors are seeking partners to further develop technology for commercialization.
Patent
N/A
Publication
Wang W., Gang G., and Stayman J., A CT denoising neural network with image properties paramterization and control, SPIE, 2021 (https://www.spiedigitallibrary.org/conference-proceedings-of-spie/11595/115950K/A-CT-denoising-neural-network-with-image-properties-parameterization-and/10.1117/12.2582145.full?SSO=1)
