Dual-energy CBCT pre-spectral-decomposition filtering with wavelet shrinkage networks

Convolutional Neural Networks (CNNs) are reshaping sig-nal processing and computer vision by providing data-drivensolutions for inverse problems such as noise reduction. How-ever, their relationship with established signal processingmethods is sometimes unclear and its development not fullyexploiting the existing knowledge. In this paper, rather thanimproving existing CNNs with wavelet transformations asexplored earlier, we improve the wavelet shrinkage approachto noise-reduction with a data-driven solution. The resultingCNN has clear encoding, decoding and processing paths.As application, we perform noise reduction in Dual-EnergyCone-Beam CT. The obtained results were compared to aUNet-like architecture, which reveal better noise-free imageswithout aliasing artifacts. This indicates that that our archi-tecture is able to preserve well the information contained inthe images because the architecture exploits explicitly theunderlying signal representation