high performance computing on graphics processing units: hgpu.org

The domain of medical imaging is naturally moving towards methods that can represent, and account for, local uncertainties in the image data. Even so, fast and efficient solutions that take uncertainty into account are not readily available even for common problems such as gradient estimation. In this work we present a CUDA implementation of Normalized Convolution, an uncertainty-aware image processing technique, well established in the signal processing domain. Our results show that up to 100X speedups are possible, which enables full resolution CT images to be processed at interactive processing speeds, fulfilling demands of both efficiency and interactivity that exist in the medical domain.