Spectral Photon Counting

• Spectral photon counting is a measurement paradigm that registers individual photon energies to enable material decomposition and energy-resolved imaging.
• It employs diverse detector architectures and calibration techniques to ensure accurate signal separation and robust inversion despite nonidealities.
• Its applications range from advanced x‑ray imaging to optical spectroscopy, offering enhanced contrast and quantitative analysis in biomedical and material studies.

Spectral photon counting denotes measurement regimes in which individual photons are registered and sorted by energy or optical frequency rather than collapsed into a single energy-integrated signal. In x‑ray imaging, this usually means pulse-height analysis with programmable thresholds on photon-counting detectors, producing multiple energy-resolved channels that can support material decomposition, K‑edge imaging, energy weighting, and dual-energy subtraction. In optical implementations, it includes single-photon spectrometers in which spectral information is encoded by interferometric beating or by channel-specific spectral responses and then reconstructed from photon counts. Across these settings, the defining feature is that spectral information is extracted from photon counting itself, so detector physics, calibration, and inversion stability become central parts of the measurement problem (Alvarez, 2017, Zheng et al., 2022, Xu et al., 2023).

1. Physical basis and mathematical formulation

In spectral x‑ray photon counting, the standard analytical framework is the Alvarez–Macovski model, which approximates the linear attenuation coefficient as a linear combination of energy-dependent basis functions, $$\mu(\mathbf{r},E)\approx \sum_{k=1}^{K} a_k(\mathbf{r}) f_k(E)$$. Along a ray, this yields basis line integrals $A_k=\int a_k(\mathbf{r})\,d\mathbf{r}$ and expected bin counts

$$\lambda_k(\mathbf{A})=\int S_k(E)\exp\!\left[-\sum_j A_j f_j(E)\right]\,dE,$$

with $S_k(E)$ determined by the incident spectrum and the bin response. Material decomposition is the inverse problem: estimate $\mathbf{A}$ from multibin data, typically after a log transform or through direct maximum-likelihood estimation (Alvarez, 2017).

At mammographic energies, a two-basis description is often used. For most natural body constituents at mammographic x‑ray energies, the attenuation model ignores absorption edges and represents attenuation as the sum of photoelectric absorption and scattering processes. In the spectral tomosynthesis prototype based on the Philips MicroDose platform, two-bin measurements are modeled as

$$m_i=\int q_0(E)\,T_i(E)\,\exp\!\big[-\mu_1(E)d_1-\mu_2(E)d_2\big]\,dE,$$

where $q_0(E)$ is the known incident spectrum, $T_i(E)$ are bin responses, and $(d_1,d_2)$ are basis thicknesses. With two independent energy measurements and two unknown path lengths, the model yields a unique solution under its assumptions (Cederstrom et al., 2021).

In spectral CT, the same principle is written in material-fraction form. For basis materials $k=1,\dots,M$ with attenuation curves $A_k=\int a_k(\mathbf{r})\,d\mathbf{r}$0 and voxel fractions $A_k=\int a_k(\mathbf{r})\,d\mathbf{r}$1, the binwise Beer–Lambert model can be expressed as

$A_k=\int a_k(\mathbf{r})\,d\mathbf{r}$2

with $A_k=\int a_k(\mathbf{r})\,d\mathbf{r}$3 the line integral of material $A_k=\int a_k(\mathbf{r})\,d\mathbf{r}$4 along ray $A_k=\int a_k(\mathbf{r})\,d\mathbf{r}$5. In the direct iterative reconstruction framework with water, iodine, gadolinium, and calcium, this polychromatic physics is summarized by a calibrated $A_k=\int a_k(\mathbf{r})\,d\mathbf{r}$6 mixing matrix and incorporated into a joint reconstruction of basis-material volume fractions (Rahman et al., 2023).

Optical spectral photon counting uses different forward models but the same logic. In photon-counting dual-comb spectroscopy, the time histogram of single-photon arrivals reconstructs an interferogram whose Fourier components map to optical comb modes. In the metasurface–SNSPD reconstructive spectrometer, the count vector $A_k=\int a_k(\mathbf{r})\,d\mathbf{r}$7 satisfies a linear model $A_k=\int a_k(\mathbf{r})\,d\mathbf{r}$8, where each channel has a distinct broadband spectral response and the unknown spectrum $A_k=\int a_k(\mathbf{r})\,d\mathbf{r}$9 is recovered computationally (Xu et al., 2023, Zheng et al., 2022).

2. Detector architectures and spectral encoding mechanisms

Spectral photon counting is implemented with several distinct detector classes. In breast imaging, the Philips MicroDose lineage uses silicon strip photon-counting detectors in a scanned-slot geometry. Each detected x‑ray photon generates an electronic pulse with amplitude proportional to photon energy; a low-energy threshold rejects noise and a high-energy threshold splits the signal into two bins. The front-end chain includes a preamplifier, shaper, discriminator(s), anti-coincidence circuitry, and counters. In the spectral tomosynthesis and spectral mammography systems, this produces a single-exposure dual-energy acquisition with sum and high-energy bins (Cederstrom et al., 2021, Fredenberg et al., 2021).

Other x‑ray implementations use CdZnTe or CdTe pixel detectors. A Medipix3RX CdZnTe system with five energy bins was used for direct iterative reconstruction of water, iodine, gadolinium, and calcium, while a 1 mm CdTe Medipix detector with $$\lambda_k(\mathbf{A})=\int S_k(E)\exp\!\left[-\sum_j A_j f_j(E)\right]\,dE,$$0 pitch was used in spectral micro‑CT together with charge-summing mode and empirical spectral correction. These architectures provide more than two energy channels, but they also amplify detector-response complexity, including pixel-dependent efficiency variation and charge-sharing effects (Rahman et al., 2023, Luna et al., 2023).

Monte Carlo assessment shows that “spectral performance” cannot be reduced to threshold count efficiency alone. In direct-converting CdTe detectors, charge sharing and K‑fluorescence produce a broadened photopeak, satellite peaks, and low-energy continuum; in edge-on silicon detectors with tungsten foils, Compton interactions dominate the high-energy response and make low thresholds important; and in an optical counting system based on LaBr$$\lambda_k(\mathbf{A})=\int S_k(E)\exp\!\left[-\sum_j A_j f_j(E)\right]\,dE,$$1 and SiPMs, spectral distortion is dominated by optical counting statistics rather than charge sharing. The same study also considers a CdTe photon counter with additional coincidence counters, designed to identify simultaneous over-threshold events in adjacent pixels and thereby recover spectral information otherwise lost to double counting (Stierstorfer et al., 2024).

Spectral encoding is not confined to x‑rays. In the reconstructive spectrometer on silicon-on-insulator, each metasurface region imposes a distinct, spectrally structured response on a colocated NbN superconducting nanowire single-photon detector. The resulting broadband but low-correlation channel responses form a measurement matrix for computational inversion. In near-ultraviolet photon-counting dual-comb spectroscopy, spectral encoding is interferometric rather than spatial: each radio-frequency comb tooth corresponds uniquely to an optical comb mode, so a single photo-counter can record multiplexed spectral data (Zheng et al., 2022, Xu et al., 2023).

3. Calibration, material decomposition, and reconstruction strategies

A recurrent feature of spectral photon counting is that inversion is calibration-sensitive. In spectral photon-counting tomosynthesis for lesion characterization, calibration proceeds in stages. Raw counts in each bin and detector channel are mapped to equivalent PMMA thickness using look-up tables from a PMMA step wedge. Each bin is then reconstructed separately into 3D stacks using a modified SART algorithm “modified to yield a globally linear behavior,” after which voxelwise dual-energy decomposition into equivalent aluminum and polyethylene thicknesses is performed using an Al–PE step-wedge calibration. In the breast-density system, a related calibration maps dual-energy measurements to equivalent Al and PE thicknesses and then to equivalent adipose and glandular thicknesses through a linear transfer based on known attenuation data (Cederstrom et al., 2021, Fredenberg et al., 2021).

Image-domain and projection-domain decompositions coexist in the literature. ROI-wise material decomposition in spectral photon-counting CT reconstructs a multienergy image for each bin, models the voxel attenuation vector as $$\lambda_k(\mathbf{A})=\int S_k(E)\exp\!\left[-\sum_j A_j f_j(E)\right]\,dE,$$2, segments the multienergy image into spatio-spectral regions of interest using Gaussian-mixture and kernel k-means procedures, and then performs sparse decomposition with an ROI-specific basis subset. The stated motivation is conditioning: global voxelwise decomposition with a fixed basis remains ill-conditioned when materials have similar attenuation curves and detector nonidealities degrade effective spectral separation (Xie et al., 2019).

A more integrated alternative is direct reconstruction of basis images. In the MBIR framework for a photon-counting spectral CT phantom, a calibrated $$\lambda_k(\mathbf{A})=\int S_k(E)\exp\!\left[-\sum_j A_j f_j(E)\right]\,dE,$$3 mixing matrix is inserted directly into the forward model, and the reconstructed variables are material volume fractions subject to non-negativity and a unity-or-less sum constraint. The objective is convex because both the data term and the regularizer are quadratic and the constraints are linear; optimization is performed by iterative coordinate descent with a modified simplex solver for each voxel (Rahman et al., 2023).

Calibration can also be empirical and per pixel. In high-resolution spectral micro‑CT, signal-to-thickness calibration with PMMA slabs is performed for each energy window and each detector pixel, fitting $$\lambda_k(\mathbf{A})=\int S_k(E)\exp\!\left[-\sum_j A_j f_j(E)\right]\,dE,$$4. Measured counts are then converted to equivalent PMMA thickness,

$$\lambda_k(\mathbf{A})=\int S_k(E)\exp\!\left[-\sum_j A_j f_j(E)\right]\,dE,$$5

and from there to corrected mass attenuation coefficients. This spectral correction is combined with Iterative Clustering Material Decomposition, in which Gaussian-mixture clustering in multienergy attenuation space restricts the candidate basis materials before local least-squares decomposition (Luna et al., 2023).

More recent work turns the decomposition step itself into a differentiable module. In end-to-end differentiable photon-counting CT, maximum-likelihood material decomposition is treated as an implicit layer. If $$\lambda_k(\mathbf{A})=\int S_k(E)\exp\!\left[-\sum_j A_j f_j(E)\right]\,dE,$$6 denotes the MLE solution and $$\lambda_k(\mathbf{A})=\int S_k(E)\exp\!\left[-\sum_j A_j f_j(E)\right]\,dE,$$7 any upstream model parameters, then

$$\lambda_k(\mathbf{A})=\int S_k(E)\exp\!\left[-\sum_j A_j f_j(E)\right]\,dE,$$8

which allows image-domain losses on reconstructed quantitative images to backpropagate into detector-threshold models or scatter-correction networks. This suggests a shift from manual, stagewise calibration toward cross-domain optimization based on the quantitative images themselves (Wang et al., 12 Feb 2026).

4. Nonidealities, conditioning, and task-based optimization

Spectral photon counting is often presented as if energy binning alone ensured stable material decomposition. The literature shows the opposite. In a pileup-inclusive analysis of the Alvarez–Macovski inverse problem, three-bin pulse-height analysis with high pileup can become ill-conditioned for specific values of object attenuation, producing Cramér–Rao lower-bound variances larger by approximately $$\lambda_k(\mathbf{A})=\int S_k(E)\exp\!\left[-\sum_j A_j f_j(E)\right]\,dE,$$9 than comparable four-bin or low-pileup configurations. The Jacobian condition number peaks on an approximately planar locus in A-space, with the ill-conditioned region associated with $S_k(E)$0. This directly contradicts the common shorthand that “three bins for three basis functions is sufficient” once realistic detector physics is included (Alvarez, 2017).

Threshold optimization is likewise task- and detector-dependent. Robustness analysis for photon-counting spectral CT shows that, when pileup and imperfect energy response are included, optimal thresholds do not always increase with increasing attenuation. Optimizing thresholds for a 30 cm phantom yields near-optimal SDNR$S_k(E)$1 or SNR$S_k(E)$2 across target materials and surrounding attenuations for both silicon-strip and CZT detectors, and using more than three bins reduces the need to retune thresholds for different anatomies. The stated design implication is pragmatic: around six or eight bins may deliver near-optimal performance without generating an unnecessarily large amount of data (Zheng et al., 2018).

Scatter is another central nonideality, and in spectral systems it is explicitly energy-dependent. For contrast-enhanced spectral mammography with a photon-counting detector, uncorrected scatter causes projected iodine densities to be increasingly underestimated as object thickness increases. An energy-sensitive scatter model based on PMMA reference scatter fractions and binwise correction improves iodine density estimation and is stated to be transferable to tomosynthesis and CT without changing acquisition parameters. This directly opposes the simplification that scatter can be treated as an energy-independent background in energy-resolved photon counting (Lewis et al., 2022).

Task-based metrics refine these conclusions. In contrast-enhanced spectral mammography with a silicon strip detector, optimizing with respect to signal-to-quantum-noise ratio alone yields only a minute improvement, whereas weights close to dual-energy subtraction, which approximately minimize anatomical noise, improve ideal-observer detectability close to 50% in the phantom study and much more strongly in dense-breast simulations. The broader implication, stated explicitly, is that inclusion of anatomical noise is essential for optimizing spectral imaging systems (Fredenberg et al., 2021).

Detector evaluation therefore extends beyond channelwise DQE. A Monte Carlo framework for four photon-counting detector types derives spatial-spectral statistics, MTF, NPS, and task-based spectral DQE directly from single-photon response simulations. In that analysis, CdTe excels for standard quantification tasks, silicon can outperform CdTe for spectral decomposition at low spatial frequencies, coincidence counters can recover much of the spectral performance lost to charge sharing, and a LaBr$S_k(E)$3 optical counting system yields the highest spectral DQE at low to moderate spatial frequencies among the architectures considered (Stierstorfer et al., 2024).

5. Biomedical applications in x‑ray imaging

Breast imaging is a prominent application domain because photon-counting detectors and mammographic energies make dual-bin acquisition practical. In lesion characterization with spectral photon-counting tomosynthesis, a prototype scanned-slot DBT system derived from Philips MicroDose was used to distinguish “cyst-like” from “tumor-like” targets. In the phantom, two targets were designed to appear similar in 2D spectral projection, but spectral tomosynthesis revealed the exact compositional difference in depth. The measured difference between the left and right target in the lower slice was approximately $S_k(E)$4 Al, compared with an expected $S_k(E)$5, demonstrating partial recovery of the compositional difference and simultaneous degradation by the limited-angle point-spread function. The paper concludes that lesion characterization is feasible in spectral tomosynthesis, but more data and refinement of calibration and discrimination algorithms are needed (Cederstrom et al., 2021).

The same detector family has also been used for quantitative volumetric breast-density estimation. In first clinical results with a spectral photon-counting tomosynthesis prototype, the precision was determined to 2.4 percentage points of volumetric breast density in phantom measurements. In a clinical cohort of 18 patients, strong correlations were observed between contralateral ($S_k(E)$6) and ipsilateral ($S_k(E)$7) density measurements, and the density distribution over 68 images was skewed in the range $S_k(E)$8 with a median of 18%. The method is explicitly positioned as more physically grounded than area-based density measures because it estimates adipose and glandular thicknesses rather than dense area alone (Fredenberg et al., 2021).

Contrast-enhanced spectral mammography with photon counting targets iodine rather than bulk composition. Using a dual-bin silicon-strip detector with measured thresholds at 20.6 keV and 32.3 keV, task-based optimization showed that optimal combination of the energy-resolved images corresponded approximately to minimization of anatomical noise, commonly referred to as energy subtraction. Under those conditions, the ideal-observer detectability index improved close to 50% in the phantom study; in a clinically more realistic simulation, spectral imaging was predicted to perform approximately 30% better than absorption imaging for an average glandularity breast and, for dense breast tissue with high anatomical noise, by a factor of 6, with another approximately 70–90% improvement stated to be within reach for an optimized system (Fredenberg et al., 2021).

Spectral photon counting CT extends these ideas to explicitly multi-material tasks. Direct iterative reconstruction of water, iodine, gadolinium, and calcium volume fractions from five-bin photon-counting data showed that volume fractions within ROIs in a low-concentration scan were close to ground truth. Reported examples include iodine errors of 18.8% and 28.4% for higher and lower iodine concentrations, gadolinium errors of 8.8% and 27.5%, and calcium errors of 20.9% and 40.9%, with water fraction errors near 1% in the water ROI. The same general problem appears in micro‑CT, where empirical spectral correction plus iterative clustering enabled quantitative separation of several materials in a phantom and separation of three distinct tissue types in mouse: muscle, fat, and bone (Rahman et al., 2023, Luna et al., 2023).

Phase-sensitive mammography introduces a related but distinct application. Photon-counting spectral phase-contrast mammography based on Talbot interferometry showed that differential phase contrast improved detectability compared to absorption contrast, particularly for fine tumor structures, and that optimal incident energy is higher in differential phase contrast than in absorption contrast. However, the paper states equally clearly that spectral material decomposition was not facilitated by phase contrast, because phase does not add an independent interaction basis beyond the density-related information already correlated with Compton attenuation (Fredenberg et al., 2021).

6. Broader modalities, misconceptions, and research directions

Beyond medical x‑ray imaging, spectral photon counting also appears in integrated and interferometric optical spectroscopy. A metasurface–SNSPD reconstructive spectrometer fabricated on SOI demonstrated reconstruction of monochromatic light with 2 nm resolution over 1500–1600 nm and total detection efficiency of 1.4%–3.2% in that band. For a broadband Gaussian spectrum, the experimentally determined condition $S_k(E)$9 was satisfied at about 30 ms measurement time, indicating that tens-of-milliseconds reconstruction is possible in a true photon-counting regime when the count vector is inverted against calibrated channel responses (Zheng et al., 2022).

Photon-counting dual-comb spectroscopy extends the same logic to the near-ultraviolet and visible. There, individual detection events are time-stamped, accumulated into an interferogram, and Fourier transformed into a spectrum whose optical frequencies are calibrated by the comb relation and the radio-frequency reference. The measured signal-to-noise ratio follows the quantum-limit scaling expected from Poisson counting statistics, and experiments were demonstrated with power per comb line as low as a femtowatt. This suggests that multiplexed spectral photon counting can remain quantitatively reliable even when analog detection would be dominated by technical noise (Xu et al., 2023).

The phrase “spectral photon counting” also appears in theoretical work on photon-emission statistics under spectral diffusion. For a driven two-level system subject to nonstationary Ornstein–Uhlenbeck or random telegraph noise, the generating-function and stochastic-Liouville analysis shows that line shape and Mandel’s parameter depend on nonequilibrium environmental characteristics at short times in the slow-modulation limit, but become independent of those characteristics in steady state; in the fast-modulation limit, neither the line shape nor the Mandel parameter depends on the nonequilibrium characteristics. A plausible implication is that time- and frequency-resolved photon-counting experiments can distinguish equilibrium from nonequilibrium environments only when spectral diffusion is slow enough for its transient asymmetries to survive into the measurement window (Cai et al., 15 Apr 2026).

Several misconceptions recur across the field. One is that more bins automatically guarantee better decomposition; pileup and imperfect response show that invertibility can still fail (Alvarez, 2017). Another is that scatter or anatomical structure can be ignored during optimization; in mammography, doing so yields nearly trivial gains compared with task-based optimization that includes anatomical noise (Fredenberg et al., 2021, Lewis et al., 2022). A third is that phase contrast provides an extra decomposition basis; the Talbot-interferometry analysis explicitly states that it does not (Fredenberg et al., 2021). Finally, there is no single preferred inversion strategy: practical systems range from step-wedge look-up tables and image-domain decomposition to direct MBIR, clustering-assisted decomposition, deep sinogram-domain correction for monochromatic image synthesis, and end-to-end differentiable pipelines that optimize detector calibration or scatter correction from quantitative image losses (Zheng et al., 2020, Wang et al., 12 Feb 2026).

Taken together, these results depict spectral photon counting as a measurement paradigm rather than a single hardware class. Its defining advantage is access to energy- or frequency-resolved photon statistics at the detector level. Its defining difficulty is that quantitative performance depends not only on counting efficiency, but also on calibration, response stability, inversion conditioning, and task-specific optimization. The published literature therefore treats detector design, spectral encoding, reconstruction, and quantitative validation as inseparable components of the same problem (Stierstorfer et al., 2024, Zheng et al., 2018).