Signal preserving non-local noise suppression for photon-counting CT

With advancements in detector technology, photon-counting CT (PCCT) in a clinical setting is becoming a reality. Photon-counting detectors typically split an x-ray spectrum into discrete bins and then reconstruct multiple images independently. Thus, each image is reconstructed with a fraction of the number of photons of an energy-integrated scan, leading to increased noise. In order to alleviate this issue, the proposed method is designed to suppress noise by extraction of the redundant structural information allowed by multiple scans. A template image is created, which is a linear combination of CT images at each energy level, with the weights determined by the contrast-to-noise-variance ratio (CNVR). An exponential model is then used on the template image to calculate a similarity coefficient between pixels based solely on their CT values. These similarity values are stored into a similarity matrix, W. Noise is then suppressed when the image vector is multiplied by W. For evaluation of the proposed method, noise-suppressed low dose images are compared against high dose images. Results are compared for both simulation and experimental studies.