Image processing to improve quality and quantification of K-edge images from spectral CT

Spectral photon counting computed Tomography (SPCCT) is a technology that uses the same principle of X-ray scanning but with a new detector type able to count incident photons and to sort them as a function of their energy. This provides more information by discrete sampling of the transmitted spectra and enables new approach for material decomposition such as K-edge imaging, which is the specific detection of one, or even multiple, contrast agents. Since one of the main limitations of the current SPCCT prototype is the presence of ring artifacts and strong noise, which prevents from obtaining good quality images, the aim of this work was to study and implement image post-processing methods to improve the quality and quantification of K-edge images in SPCCT. The focus was mainly on the cardiovascular application consisting of heart images and phantom images simulating vessel calcification, with different degrees of stenosis that are impossible to detect and classify relying on standard CT images. For the ring artifact removal the approach was based on sinogram processing technique, by applying a spline smoothing of the projections, thus erasing the stripes that represent the artifacts in the sinogram domain. Different de-noising approaches were applied: local smoothing (Gaussian filter, combing features filter, and anisotropic diffusion filter), non- local smoothing (non-local mean filter), and frequency domain (wavelets). Among the implemented de-noising methods, the combing features filter showed the best qualitative and quantitative results. This filter gives the possibility to combine the normalized results of the different standard filters, trying to take the advantages of each one. The quantification of the concentric stenosis in the phantom K-edge images allowed to validate the implemented methods. The calculated radii of the lumen vessels, mediated on 4 slices, after de-noising and segmentation were of 0.76 mm with σ=0.06 on a true value of 0.75 mm, 1.46 mm with σ=0.12 on a true value of 1.5 mm, and 3.15 mm with σ=0.09 on a true value of 3 mm. This quantification on the same images without de-noising showed a not acceptable over estimation of the lumen vessels, demonstrating the importance of these post-processing techniques to take the advantages of the spectral detection.