Deep neural networks (DNN) are currently used in wide range of computer vision and image processing tasks. Adopting DNN-based techniques for safety-critical clinical applications, such as MRI analysis and reconstruction, in which results inform diagnostic, prognostic, and interventional decisions in medicine and scientific conclusions in research requires mechanisms to determine the reliability of the DNN predictions by obtaining an estimation of the uncertainty of the DNN in the predictions.
The NPB-DNN technology presents a non-parametric Bayesian approach to estimate the uncertainty in DNN-based algorithms. Thus, enables the utilization of DNN-based algorithms in safety-critical environments. The NPB-DNN approach provides a principled way to characterize the posterior distribution of the solution, thus it can simultaneously provide an estimation of the uncertainty in the DNN predictions and improve generalization capabilities. Further, NPB-DNN can be used as a plug-in to existing DNN-based algorithms without any modification to the DNN architecture.
- Improved DNN-based algorithms accuracy
- Robustness and high performance in noisy scenarios
- Reliable predictions for safety-critical environments
- Out-of-distribution data prediction
Applications and Opportunities
- Medical imaging: Registration/Segmentation/Accelerated MRI reconstruction/image enhancement
- Computer vision: Scene understanding/optical flow estimation/super-resolution