Image Quality Issues Specific to Spectral CT
Spectral CT scanners can reconstruct regular CT-like volumes: the photon counts obtained from a spectral CT acquisition can be either fed to one of the spectral CT reconstruction methods described above, yielding material-specific maps, or merged back together into a single sinogram and reconstructed. generating a regular-CT volume (e.g., in Hounsfield Units, HU). Although they are reconstructed from identical input data, it turns out that material-specific CT volumes are typically much noisier than their HU counterpart. The two fundamental reasons for this phenomenon are: reconstructing several volumes instead of a single one reduces the amount of measured photons used per voxel, which results in higher noise (which can be compensated for by increasing the radiation exposure), and the non-linear decomposition process amplifies the noise.
With the introduction of photon-counting detectors into clinical routine, one can expect to see a reduction in detector pixel sizes [69, 84]. The increase in spatial resolution will extend the diagnostic range of CT imaging, for example, in the visualization of fine structures in the lung or along coronary arteries with stents [37, 65, 68]. In those cases, the high-resolution acquisition enables an improved sampling of high-frequency features and reduces noise aliasing . However, for sections without fine details, a high-frequency noise will significantly reduce the image quality. In the future, it will be essential to incorporate these new circumstances into the image reconstruction and to optimize it through algorithmic solutions still to be developed. On this note, the additional energy dimension provides an increased amount of information, which can be utilized to denoise spectral images. The data can be utilized following strategies like prior image constraints  or dictionaries [39, 85] (section 19.4).
Additionally, material volumes are subject to decomposition errors, commonly referred to as “cross-talk”: materials can appear in the wrong material-specific CT volumes. The severity of cross-talk depends on how much the materials’ attenuation profiles differ from each other (the more similar they are, the stronger the cross-talk) and on how much noise is present in the photon counts (the noisier the data, the stronger the cross-talk). In one-step inversion methods, regularization can also cause cross-talk: regularizing one material creates discrepancies between the estimated photon counts and the measured ones, which are compensated by adding or removing some amount of another material. This effect is particularly intense on the borders of structures when a strong spatial regularization is applied, as illustrated in Figure 19.7.
Ring artifacts are a very common artifact in any type of CT imaging and can have a variety of sources. In conventional CT, if one detector element is out of calibration, the reading of this element may consistently be incorrect. As a consequence, the later reconstructed CT slice will be affected by rings. As photon-counting detectors are highly complex and sensitive compared to conventional detectors, a dedicated calibration needs to be performed. While this spectral technology, as well as calibration methods, are still under development, rings that may appear after reconstruction can be removed to a large degree by classical ring removal algorithms [45, 88]. Regarding rings or other artifacts, it is essential to understand that the current hardware does not represent an ideal detector. Novel sensor material (imperative for photon-counting CT), along the lines of cadmium telluride and cadmium zinc telluride, come with technical challenges which can be addressed by hardware as well as software solutions. Pile-up and spectral distortions are two of the main effects, which reduce the quality of spectral data from photon-counting detectors. Several investigators have developed techniques to model those shortcomings with different software-based techniques [10,55,70,72-74]. These achievements represent an ideal opportunity to overcome some of those hardware shortcomings but they still need to be integrated in the image formation algorithms described in this chapter.