Chapter 13: Spectral CT and Photon Counting

In Chapter 11, the polychromatic nature of X-rays was a nuisance to be corrected: beam hardening warps the image, and water correction makes it go away. This final expository chapter flips the perspective. The energy dependence of attenuation ($\mu = \mu(E)$) is not noise; it is information. Measure with energy resolution and you can answer not just "how attenuating is this pixel" but "what is this pixel made of". That is spectral CT, and its ideal form is photon-counting CT (PCCT). The second half of the chapter surveys, with references, where the reconstruction story told in this textbook stands in current research.

Attenuation is built from two physical processes

In the diagnostic energy range (roughly 40–140 keV), X-ray attenuation is essentially the sum of two interactions: the photoelectric effect, strongly dependent on atomic number $Z$ and falling off as $E^{-3}$, and Compton scattering, roughly proportional to electron density and nearly energy-independent. Alvarez and Macovski pointed out in 1976 that, as a consequence, the $\mu(E)$ of any material can be written as a linear combination of two basis functions:

$$\mu(x, y; E) \;\approx\; a_1(x,y)\, f_{\mathrm{PE}}(E) + a_2(x,y)\, f_{\mathrm{C}}(E)$$

The basis can be the physical components (photoelectric/Compton) or a material pair (water/iodine). What matters is that there are only two degrees of freedom: measure at two energies and you can solve a linear system per pixel to separate materials.

Iodine ($Z=53$), the workhorse contrast agent, is photoelectric-dominated: its $\mu(E)$ varies steeply and even has a discontinuity, the K-edge at 33.2 keV. Water (most of the human body) is nearly flat. This difference in shape is the entire basis of material discrimination.

Dual-energy CT

Vendors implement two-energy acquisition in different ways: dual-source (two tube–detector pairs mounted 90° apart), rapid kV-switching (alternating tube voltage between projections), and dual-layer detectors (splitting low- and high-energy photons in depth). Whichever the route, once you have a low-energy image $x_L$ and a high-energy image $x_H$, a 2×2 system per pixel

$$\begin{pmatrix} x_L \\ x_H \end{pmatrix} = \begin{pmatrix} \mu_w(E_L) & \mu_i(E_L) \\ \mu_w(E_H) & \mu_i(E_H) \end{pmatrix} \begin{pmatrix} c_w \\ c_i \end{pmatrix}$$

yields a water-density map $c_w$ and an iodine-concentration map $c_i$ (image-domain decomposition). From those basis images you can then synthesize an image at any energy: $\hat\mu(E) = c_w \mu_w(E) + c_i \mu_i(E)$. This is the virtual monoenergetic image (VMI). Synthesize low-keV for boosted iodine contrast, or high-keV to suppress metal artifacts and beam hardening. Clinical applications are broad: virtual non-contrast (VNC) images that spare a pre-contrast scan, distinguishing uric-acid from calcium stones, imaging urate deposits in gout, and more.

Simulation: material decomposition and VMI

A phantom of water (with density variations), bone, and three iodine inserts of different concentrations is scanned at two monoenergetic energies $E_L$ / $E_H$ and decomposed in the image domain. Check three things: sliding the VMI down to 40 keV multiplies iodine contrast roughly fourfold; the bone leaks into the "iodine" map (the limit of a two-material basis; real scanners add three-material or constrained decompositions here); and with noise on, the basis images amplify noise, worse the closer $E_L$ and $E_H$ are, because the linear system becomes ill-conditioned.

Photon-counting CT

PCCT does dual-energy's "measure twice" with a single detector, at the granularity of individual photons. A conventional energy-integrating detector (EID) converts X-rays to light in a scintillator and integrates it in a photodiode: individual photons are indistinguishable, the output is one energy-weighted sum in which low-energy photons (the contrast-rich ones) are underweighted, and electronic noise is added in. A photon-counting detector (PCD) converts each photon directly into a charge pulse in a semiconductor (CdTe or Si), compares the pulse height (≈ photon energy) against thresholds, and counts photons per energy bin.

Four advantages follow directly from the principle. (1) Electronic noise falls below the lowest threshold and is simply not counted, which is decisive at very low dose. (2) Detector elements can be made smaller, raising spatial resolution (no scintillator light spread). (3) One scan yields multi-energy data, i.e. spectral imaging without a dedicated protocol. (4) Freedom in energy weighting extracts more CNR from the same dose.

Clinical deployment is underway. The Siemens NAEOTOM Alpha received FDA clearance as the first clinical PCCT in 2021, and GE Photonova Spectra received FDA 510(k) clearance in 2026. Reviews published in 2025 show that evaluation is shifting from technical feasibility toward diagnostic performance and patient impact across cardiovascular, thoracic, abdominal, musculoskeletal, neurologic, and pediatric imaging. Differences between systems, pileup and charge-sharing correction, spectral calibration, data volume, cost, and workflow remain important limitations.

Recent reconstruction research

This textbook has traveled from analytic reconstruction (FBP) through iterative methods with regularization (TV) to learned methods (CNN/U-Net). Here is how the frontier is moving.

Diffusion-model reconstruction. Recent work learns a generative prior for clean CT images with a diffusion or score-based model, then incorporates sinogram consistency during generation. Its structure resembles ASD-POCS with a learned prior in place of TV. Diffusion methods are being studied for sparse-view, limited-angle, low-dose, and material-decomposition problems, but clinical CT reconstruction still relies primarily on FBP, iterative methods, and commercial DLR. Computation time, distribution shift, data consistency, and hallucination remain unresolved.

Implicit neural representations (INRs). An INR represents an image as a network mapping coordinates to attenuation and fits it to the measured sinogram for each case. Some variants require no external training data, but that does not make an underdetermined inverse problem unique: representation bias and regularization can still produce errors. Faster optimization, 3D representations, and integration with physics models are active topics.

Self-supervision, uncertainty, and open data. Self-supervised methods reduce the need for paired high- and low-dose scans, while other methods estimate uncertainty along with the image. Spectral datasets are also appearing, including a public cone-beam PCCT dataset. Raw projection data nevertheless remain difficult to obtain, and external validation across scanners and hospitals, reproducibility, regulation, and post-approval model updates remain barriers. Foundation-model reconstruction is exploratory rather than established clinical practice.

Summary

The main storyline ends here. Chapters 1–6 covered physics and geometry (attenuation, projection, FBP, fan/cone beam), and Chapter 7 carried the same physics into industrial CT; Chapters 8–10 the algorithmic arc (iterative methods, dose and noise, compressed sensing); Chapters 11–12 the practical problems (artifacts, deep learning); and this chapter the hardware frontier and research landscape. Chapter 14 is the integrated playground: combine phantoms, geometries, methods, and parameters freely, and experiment with everything you have learned.

References

Spectral CT / PCCT

• Alvarez RE, Macovski A. Energy-selective reconstructions in X-ray computerised tomography. Physics in Medicine & Biology 21, 733–744 (1976) — the original basis-decomposition paper.
• McCollough CH et al. Dual- and Multi-Energy CT: Principles, Technical Approaches, and Clinical Applications. Radiology 276, 637–653 (2015).
• Willemink MJ et al. Photon-counting CT: Technical Principles and Clinical Prospects. Radiology 289, 293–312 (2018).
• Schwartz FR et al. Photon-Counting CT: Technology, Current and Potential Future Clinical Applications. Radiology 314, e240662 (2025).
• van der Bie J et al. Photon-counting CT: An updated review of clinical results. European Journal of Radiology 190, 112189 (2025).
• FDA. FDA Clears First Major Imaging Device Advancement for Computed Tomography in Nearly a Decade (2021) — NAEOTOM Alpha.
• FDA. 510(k) clearance letter: Photonova Spectra (2026).

Learned reconstruction / datasets

• Song Y, Shen L, Xing L, Ermon S. Solving Inverse Problems in Medical Imaging with Score-Based Generative Models. ICLR (2022).
• Chung H, Kim J, McCann MT, Klasky ML, Ye JC. Diffusion Posterior Sampling for General Noisy Inverse Problems. ICLR (2023).
• Ravi N et al. Diffusion models for medical image reconstruction. BJR Artificial Intelligence 1, ubae013 (2024).
• Molaei A et al. Implicit Neural Representation in Medical Imaging: A Comparative Survey (2023).
• A cone-beam photon-counting CT dataset for spectral image reconstruction and deep learning. Scientific Data (2025).