Physically Interpretable Deep Learning Reconstruction for Photon Counting Spectral CT

Computed tomography (CT) imaging has been playing an important role in clinical diagnostics for more than half a century. Conventional medical CT uses energy-integrating detectors (EID) to obtain structural information of a scanned object by reconstructing its effective attenuation map under radiation from a polychromatic X-ray source. Since an energy-resolved photon counting detector (PCD) enables material discrimination, improvement in spatial resolution, and reduction in radiation doses, spectral CT with PCDs has been extensively studied and entered clinical practice recently. PCDs collect photons and record them in multiple energy bins as counts. They provide more information while leading to a more complex problem for data processing and image reconstruction compared with EID. In this chapter, we will address the ideas and research progresses in model-based deep learning methods for PCD spectral CT. We will start from the basic forward models for a PCD spectral CT, describing the fundamental mathematics of spectral CT imaging with the detector response of PCD incorporated. Many studies have been published in the literature in last decades. We will briefly review the conventional spectral CT reconstruction methods and detector calibration for PCD. The main focus of this chapter will be on the deep learning methods for spectral CT reconstruction based on a roadmap from conventional spectral CT reconstruction, which make it quite explainable. The learning can be done in different data processing steps and in multiple data domains including photon counts, sinograms, virtual monochromatic images, and decomposition coefficient images. All these learning can be integrated to form a comprehensive framework supporting end-to-end learning. Aspects on loss functions, network architecture, datasets are covered. Overall, we are intended to provide a route for incorporating deep learning in spectral CT reconstruction to readers.