Photon-Counting Technology

• Photon-counting technology is a detection paradigm that registers individual photon events, preserving spectral, timing, and photon-number information.
• It leverages varied architectures—from semiconductor direct conversion to EMCCD amplification and superconducting sensors—to optimize performance in X-ray, UV-visible, and quantum-optical regimes.
• The approach addresses challenges like noise management, dead time, and calibration while enhancing applications in imaging, spectroscopy, and metrology.

Searching arXiv for recent and foundational papers on photon-counting technology across detector families and applications. Photon-counting technology comprises detector architectures and measurement methods in which individual photons are registered as discrete events rather than absorbed into a purely energy-integrating signal. Depending on the platform, a counted event may carry photon number, arrival time, pulse height, or threshold-bin information. In X-ray systems this usually means direct conversion in a semiconductor and digital thresholding of each pulse; in visible and near-infrared systems it may mean thresholded electron multiplication, deep-sub-electron readout, or superconducting detection; in quantum-optical settings it can extend to full counting statistics, waiting-time distributions, and even nondestructive photon-number measurement (Hameed et al., 2024, Harpsøe et al., 2011, Malz et al., 2019, Kansanen et al., 2024).

1. Core concept and statistical framework

The defining distinction between photon counting and energy integration is event resolution. In a photon-counting detector for X-ray imaging, individual photons are directly converted into electrical pulses whose amplitudes are proportional to photon energy; user-defined thresholds then sort events into energy bins and increment counters (Hameed et al., 2024). In EMCCD-based photon counting, the same discrete-event logic is implemented statistically: the electron-multiplication register amplifies single photoelectrons so that thresholding or Bayesian inference can estimate the most probable flux per pixel from a stack of short exposures (Harpsøe et al., 2011). In superconducting devices, photon absorption generates quasiparticles or hotspots that are read out as individually countable pulses, again preserving the discreteness of the incident field (Jorel et al., 2011, Wang et al., 2024).

Across these implementations, the first-order counting model is Poissonian. For counted photons $N$, the variance satisfies $\sigma^2 \approx N$, so the signal-to-noise ratio scales as $\sqrt{N}$ (Hameed et al., 2024). In low-flux EMCCD thresholding, the ideal per-pixel flux estimator follows

$\hat{\lambda} = -\ln(1-f),$

where $f$ is the fraction of frames above threshold (Harpsøe et al., 2011). In time-of-flight systems, photon arrival times are converted to range by

$z = \frac{c t}{2},$

so timing precision directly determines depth precision (Howland et al., 2013, Wang et al., 2024, Li et al., 2021).

This statistical discreteness is not merely a low-light convenience. It is what enables spectral discrimination, coincidence analysis, high-order correlation measurements, and task-dependent weighting strategies that are inaccessible to conventional integrating detectors.

2. Detector architectures and material platforms

Photon-counting technology is not a single detector class but a family of architectures optimized for different spectral bands, flux regimes, and observables.

Platform	Spectral regime	Characteristic capability
CZT/CdTe/CdZnTe and related X-ray PCDs	X-ray	Direct conversion, threshold bins, high count-rate imaging
EMCCD, QIS, deep-sub-electron CMOS	UV-visible-NIR	Thresholded or quantized low-light imaging
STJ, SNSPD, SMSPD	Near-IR to X-ray	Single-photon timing and, in some cases, energy discrimination
TES and MKID	Visible to SWIR and HEP regimes	Calorimetric or resonant photon-number/spectral measurement

Semiconductor X-ray photon counters are typically based on high-$Z$ materials such as CdTe, CdZnTe, GaAs, or CZT. A representative room-temperature CZT module integrates detector, high-voltage supply, ASIC readout, and digital base board, supports up to five non-overlapping energy bins, offers a nominal energy range of 20–160 keV, and uses threshold registers related to photon energy by $$\mathrm{Energy\,(keV)} = 1.33 \times \mathrm{Threshold} + 4.698$$ (Hameed et al., 2024). Fine-pitch CdTe systems such as PIXIRAD implement two simultaneous thresholds per pixel and therefore produce two “color” images per exposure in real time (Bellazzini et al., 2012). Hybrid photon counting detectors extend the same principle through separately optimized sensor and CMOS readout layers, with high quantum efficiency, ultra-low dark counts, and count-rate capability $>10^{6}$ counts per pixel per second (Fonseca et al., 11 Feb 2026).

Visible-band and UV photon counting has followed two distinct paths. One path uses gain: EMCCDs amplify single photoelectrons in a stochastic multiplication register, then apply thresholding or Bayesian inference (Harpsøe et al., 2011, Ludwick, 2022). The other path reduces read noise so far that multiplication becomes unnecessary. Quanta Image Sensors and deep-sub-electron CMOS cameras exemplify this second path: a $1024 \times 1024$ color QIS with 1.1 $\sigma^2 \approx N$0m pitch was reported with both single-bit and multi-bit photon counting (Gnanasambandam et al., 2019), while the C-BLUE 3 PC visible CMOS camera reported photon-counting mode read noise around $\sigma^2 \approx N$1 median and $\sigma^2 \approx N$2 rms, enabling discrete 0, 1, 2, … photoelectron peaks without on-chip amplification (Gach et al., 2022).

Superconducting devices occupy a third branch. Ta/Al–AlOx–Al/Ta superconducting tunnel junctions provide photon counting with chromaticity from the near infrared to X-rays, with quantum efficiency up to about 80% and resolving power exceeding 10 in the visible (Jorel et al., 2011). NbN superconducting microstrip single-photon detectors provide large-area free-space coupling and low-jitter detection at 1550 nm (Wang et al., 2024). At a broader systems level, the VUV-to-SWIR detector landscape includes SNSPDs, MKIDs, TES arrays, SiPMs, Skipper-CCDs, and Ge-on-Si SPADs, each trading off quantum efficiency, dark count rate, timing jitter, operating temperature, and scalability (Asaadi et al., 2022).

3. Spectral discrimination, timing, and number resolution

A central property of photon counting is that it can preserve more than intensity. X-ray implementations preserve approximate energy by pulse-height thresholding. In the MXA-128 CZT study, up to five non-overlapping bins were defined and used to probe soft-tissue transmission at 21, 25, 29, 31, and 45 keV (Hameed et al., 2024). PIXIRAD operationalizes the same principle in a simpler dual-threshold form: one counter records all photons above the low threshold, another all photons above the high threshold, and the difference image isolates the low-energy contribution (Bellazzini et al., 2012).

In spectral mammography, the combination of photon-counting detectors with Talbot interferometry creates spectral differential phase-contrast imaging. There, the physics of absorption and phase lead to different optimal spectral weights: the review grounded in Fredenberg and collaborators reports $\sigma^2 \approx N$3 for absorption contrast and $\sigma^2 \approx N$4 for phase contrast, with the interferometer design energy imposing an additional maximum near $\sigma^2 \approx N$5 (Fredenberg et al., 2021). This does not imply that phase contrast automatically improves every material task; the same analysis states that spectral material decomposition was not facilitated by phase contrast, even though fine tumor structures benefited in detectability (Fredenberg et al., 2021).

Photon counting can also preserve number resolution directly. In Ta STJs, the integrated pulse charge scales with photon energy, and the measured spectrum at $\sigma^2 \approx N$6m exhibited regularly spaced peaks for 1, 2, … photons. The mean collected charge per photon was reported as $\sigma^2 \approx N$7, with an estimated tunneling multiplication $\sigma^2 \approx N$8 (Jorel et al., 2011). In visible CMOS without avalanche gain, multi-level thresholding between histogram valleys similarly produces “quanta images” of integer photon counts (Gach et al., 2022).

Timing is the parallel axis of information preservation. Time-magnified TCSPC implemented a temporal imaging system with magnification

$\sigma^2 \approx N$9

using $\sqrt{N}$0 and $\sqrt{N}$1, and experimentally achieved an effective single-photon timing resolution of 550 fs with an off-the-shelf SPAD (Li et al., 2021). At a less extreme but directly deployable level, a free-space-coupled superconducting microstrip detector operated at 1550 nm with system jitter of about 171 ps at an imaging operating point of $\sqrt{N}$2 and $\sqrt{N}$3 (Wang et al., 2024).

A broader theoretical generalization appears in microwave quantum optics, where photon counting is treated through full counting statistics. In Gaussian bosonic networks, introducing counting fields transforms the Lyapunov dynamics of the covariance matrix into a matrix Riccati equation, from which emitted and absorbed photon statistics, waiting-time distributions, and second-order coherence functions can be calculated exactly (Kansanen et al., 2024). This suggests that photon counting is not only a detector technology but also a statistical measurement framework.

4. Imaging, spectroscopy, and metrology applications

In medical and preclinical X-ray imaging, photon counting is used to recover spectral contrast, improve spatial resolution, and reduce electronic-noise penalties. The CZT soft-tissue study found an optimal frame rate of 12 FPS for 35 kVp/1.0 mA operation near the spectral peak, reported linear count-rate behavior up to $\sqrt{N}$4 counts/s/pixel, and observed tissue-dependent count differences across thickness and threshold (Hameed et al., 2024). A broader PCCT review reports reduced noise, higher gray-white matter CNR in brain imaging, improved interstitial lung and musculoskeletal detail in ultra-high-resolution modes, and extensive use of VMIs and QIR in clinical pipelines (Alves et al., 2024). A plausible implication is that photon counting has shifted CT from a primarily morphologic modality toward a jointly morphologic and spectral one.

Phase-sensitive mammography represents a more specialized application. In grating-based spectral differential phase-contrast mammography, detectability improved relative to absorption contrast particularly for fine tumor structures, while optimal incident energy was higher in differential phase contrast than in absorption contrast (Fredenberg et al., 2021). This is an example of photon counting being valuable not because every photon is counted per se, but because the detector preserves the spectral structure needed for task-specific weighting.

Laboratory X-ray tomography has also been reshaped by large-area hybrid photon counting detectors. A 4-megapixel hybrid photon counting detector was used for laboratory nano-xCT of an integrated circuit fabricated at the 130-nm node, collecting 40 times more photons 20 times faster than the group’s previous work, for an overall speedup of $\sqrt{N}$5. The reported reconstruction achieved 75–80 nm spatial resolution on 160 nm wiring features, validated with MTF, FSC, and CNR (Fonseca et al., 11 Feb 2026).

Photon-counting lidar and ToF imaging form another major branch. A compressive single-pixel lidar system combined time-correlated single-photon counting with random Hadamard projections to reconstruct both intensity and depth from a single under-sampled measurement set, reaching $\sqrt{N}$6 transverse resolution with acquisition times as short as 3 s, and $\sqrt{N}$7 real-time video at 14 frames per second (Howland et al., 2013). The SMSPD study demonstrated free-space ToF imaging at 1550 nm with a 0.5 m stand-off and reconstructed intensity maps with signal-to-background ratio above 80 at bright points (Wang et al., 2024). Time-magnified TCSPC then pushed timing precision to the point where range-walk error was suppressed by 99.2%, yielding depth measurement accuracy of $\sqrt{N}$8m and precision of $\sqrt{N}$9m (Li et al., 2021).

Photon counting also appears in domains where the measured object is not an image. In Bragg spectroscopy of ultracold gases, heterodyne photon counting measures depletion or amplification of a weak Bragg beam, achieving shot-noise-limited detection of the optical response rather than relying on time-of-flight atom imaging (Pino et al., 2010). In quantum optics, an SNSPD array plus FPGA-based TDC/ACDC system measured coherent-state statistics using up to 8th-order correlations and reconstructed coherent-state photon-number distributions up to average photon number about 4 (Lusardi et al., 2018). In waveguide QED, the counting problem can even be mapped onto atom counting by shelving collective excitations into a metastable atomic state, with a nondestructive variant possible for symmetric Dicke-state preparation (Malz et al., 2019).

5. Noise sources, calibration, and nonideal response

The practical difficulty of photon counting lies less in detecting a single event than in discriminating real events from detector-induced events at scale. In semiconductor X-ray PCDs, the dominant nonidealities include dead time, pulse pile-up, charge sharing, charge loss, and threshold drift. The CZT characterization study explicitly notes that no corrections for dead time, pulse pile-up, charge sharing, threshold drift, or charge loss were discussed, and that detector reliability below about 20 keV was affected by electronic noise (Hameed et al., 2024). Fine-pitch CdTe sensors face the additional issue that charge sharing near pixel boundaries can distort threshold-bin assignment (Bellazzini et al., 2012).

EMCCD photon counting replaces one set of problems with another. The principal backgrounds are clock-induced charge, dark current, and threshold errors. The Bayesian EMCCD study identifies parallel CIC as indistinguishable from photoelectrons and serial CIC as a distinct lower-gain exponential component; in the tested camera, the background floor seen by both methods was about 0.045 electrons/pixel/read, dominated by pCIC plus sCIC (Harpsøe et al., 2011). Thresholding remains close to shot-noise limited up to roughly one photon per pixel per readout, but thresholding deteriorates above $\hat{\lambda} = -\ln(1-f),$0–1.5 because of coincidence loss, while Bayesian inference remains linear to higher flux before approaching the $\hat{\lambda} = -\ln(1-f),$1 excess-noise regime (Harpsøe et al., 2011). The analytic SNR treatment developed for Roman Telescope planning further shows how threshold choice, CIC, dark current, and photometric correction all enter the final photon-counting performance budget (Ludwick, 2022).

A distinct calibration problem arises in UV silicon detectors. The JPL Microdevices Lab study used delta-doped photon-counting EMCCDs to investigate the dark-current plateau at low temperature and found that the apparent dark rate was strongly affected by the ambient environment. Lowering the shroud from about 293 K to about 230 K reduced the measured dark rate from 0.453 to 0.346 pix$\hat{\lambda} = -\ln(1-f),$2 hr$\hat{\lambda} = -\ln(1-f),$3 at 183 K detector temperature and from 0.0674 to 0.0320 pix$\hat{\lambda} = -\ln(1-f),$4 hr$\hat{\lambda} = -\ln(1-f),$5 at 173 K, suggesting that ambient blackbody photons or light leaks contribute materially to the low-temperature floor (Khan et al., 2024). This directly challenges the common assumption that a low-temperature plateau is purely intrinsic silicon dark current.

Superconducting systems exhibit a different trade space. In the SMSPD, increasing bias current improves timing behavior and internal detection efficiency but also increases dark count rate: free-space SDE reached about 5.2% at 1550 nm with DCR around 200 kcps at $\hat{\lambda} = -\ln(1-f),$6A, while the imaging operating point of $\hat{\lambda} = -\ln(1-f),$7A was chosen to balance $\hat{\lambda} = -\ln(1-f),$8, $\hat{\lambda} = -\ln(1-f),$9, and 171 ps jitter (Wang et al., 2024). In ToF systems, pile-up can manifest not just as missed counts but as range-walk error. In the time-magnified TCSPC study, increasing detection probability from 0.11% to 1.04% shifted the timing peak by about 15 ps in conventional operation; the time-magnified architecture suppressed that error by 130 times (Li et al., 2021).

These cases show that photon counting is never simply “noise-free.” Rather, it reorganizes the noise problem around threshold setting, pulse-shape fidelity, environmental backgrounds, and count-rate-dependent distortions.

6. Historical development, present constraints, and research directions

Historically, photon counting in the visible began with image intensifiers and centroiding systems developed by Boksenberg and collaborators in 1972, then moved through EMCCDs in the 2000s, and has now expanded toward deep-sub-electron CMOS and quanta-imaging architectures (Gach et al., 2022). In X-ray science, hybrid photon counting detectors have matured over the last 30 years at synchrotrons and are now being adapted to laboratory tomography (Fonseca et al., 11 Feb 2026). In HEP and related low-background experiments, the detector portfolio now spans Skipper-CCDs, SiPMs, Ge-on-Si SPADs, SNSPDs, MKIDs, and TES arrays from the vacuum ultraviolet to the short-wave infrared (Asaadi et al., 2022).

Several current constraints recur across these lines of development. In clinical and preclinical PCCT, remaining needs include quantifying energy resolution, managing high-flux effects such as pile-up and dead time, calibrating thresholds across temperature and bias, modeling charge sharing, and optimizing bin settings for specific clinical tasks (Hameed et al., 2024, Alves et al., 2024). In EMCCD photon counting, lower-CIC electronics, improved controllers, and accelerated Bayesian processing remain active targets (Harpsøe et al., 2011). In superconducting ToF detection, lower operating temperature, better out-of-band filtering, improved optical cavities, and cryogenic amplifiers are explicit routes toward lower DCR and smaller jitter (Wang et al., 2024). In UV and HEP instrumentation, direct VUV detection without wavelength shifters, SWIR-capable arrays with low dark counts, and cryogenic ASIC integration are identified as major technology opportunities (Asaadi et al., 2022).

One point of ongoing methodological tension concerns amplification versus read-noise reduction. EMCCDs and e-APDs demonstrate that pre-readout amplification can recover single-photon sensitivity, but visible-band CMOS work argues that excess noise factor prevents accurate multiple photon counting in amplified detectors and that the only solution is to lower the readout noise (Gach et al., 2022). This is not a universal theorem across all spectral bands; rather, it marks a platform-dependent boundary between thresholded single-photon detection and true proportional photon-number resolution.

At the conceptual edge of the field, photon counting is also expanding beyond destructive detection. Waveguide-QED schemes propose reducing photon counting to atom counting with fidelity that improves with emitter number, including a nondestructive photon-number measurement in which the incident state is scattered into the waveguide unchanged (Malz et al., 2019). In Gaussian bosonic networks, the same observable is treated through exact counting-field formalisms that connect cumulants, coherence, entanglement witnesses, and nonreciprocity (Kansanen et al., 2024). This suggests that future photon-counting technology will likely converge detector engineering, statistical inference, and quantum measurement theory more tightly than in earlier generations.

Photon counting therefore denotes both a hardware paradigm and a measurement philosophy: preserve the discreteness of the radiation field long enough that energy, time, number, and correlation structure remain experimentally actionable. Across X-ray CT, UV astronomy, visible low-light imaging, SWIR sensing, ToF metrology, quantum optics, and HEP instrumentation, that principle has become technically mature enough to support large-area systems, spectral analysis, deep-learning pipelines, and sub-picosecond timing, while still leaving open fundamental problems in calibration, count-rate linearity, and true multi-photon number resolution.