Key words
Photon Counting - Computed Tomography - Diagnostic Imaging - Spectral Computed Tomography
- Photon-Counting Detector - Energy-Integrating Detectors
Introduction
Photon-counting computed tomography (PCCT) was launched in the clinical routine in
July 2021 and has the potential to fundamentally change today’s workflows. This article
summarizes the technical principles and potential clinical benefits of the new photon-counting
detector (PCD) technology in comparison to the currently established energy-integrating
detectors (EID) and presents initial experiences from daily practice.
Since its clinical introduction in the 1970 s, computed tomography has revolutionized
diagnostic workflows and patient management [1]. The developments from the early beginnings to current clinical practice were only
possible due to major innovations such as the development of spiral CT and multi-slice
systems with a higher pitch, as well as the acquisition of dual energy (DE) data at
a comparable radiation dose to that of single-source CT systems. This has significantly
expanded the range of indications to include vascular, cardiac, and functional imaging
[2]
[3]
[4]
[5].
However, a noteworthy disadvantage of these EID systems can be attributed to the fact
that the energy of the X-ray photons is measured indirectly using a scintillator that
transforms incoming X-ray photons into visible light before the signal is converted
into an electrical current ([Fig. 1a]). Moreover, the detected signal is a cumulative measure of the energy of all incoming
X-ray photons and does not contain information about their absolute number and individual
energy level.
Fig. 1 a Energy-integrating scintillation detector. Individual detector cells made of a scintillator
such as gadolinium oxide or gadolinium oxysulfide (GOS) absorb the X-ray photons and
convert their energy into visible light. Subsequently, this light is detected by photodiodes
placed on the back of each detector cell and converted into electric current. The
detected signal is proportional to the cumulative photon energy without the ability
to differentiate between individual photons. b In a PCCT detector, a semiconductor such as cadmium telluride absorbs the X-ray photons.
This creates electron-hole pairs in a number proportional to the detected photon energy,
which are separated into a strong electric field, resulting in a direct conversion
of the detected signal into electric current. With this approach, individual photons
can be counted and their respective energy can be measured.
The photon-counting technology, which has been developed over the last two decades,
has the potential to overcome these limitations by solving several physical challenges,
such as signal splitting at the borders of the detector pixels, energy loss of X-rays
due to K-escape, and the so-called “pulse pile-up”, where the high photon flux density
commonly seen in medical CT can cause overlapping low-energy pulses to be falsely
registered as high-energy hits [6]
[7]. After successful implementation in pre-clinical scanners, photon-counting detector
CT has now entered the clinical routine and reduces the limitations of energy-integrating
CT as elucidated in the remainder of this review. While the first clinical system
was introduced by Siemens Healthineers in late 2021, other major vendors such as Philips,
GE, and Canon are expected to enter the market with clinical photon-counting CT systems
soon.
Photon-counting detector – technical background
Photon-counting detector – technical background
Current EID CT relies on solid-state scintillation detectors consisting of a scintillator
with septa and a photodiode array. Detector cells have a width of about 0.25–0.625 mm,
projected to the CT’s isocenter [8]. Incoming X-ray photons generate a shower of visible light within the scintillator,
which is subsequently converted into an electric signal by the photodiode. Here, the
detected signal is proportional to the total energy of all photons during a measurement
interval, without specific information about an individual photon or its energy. The
septa in the scintillator material are necessary to partition the different detector
elements. This prevents light photons from crossing between the detector elements.
However, they limit the geometric dose efficiency, especially when trying to build
smaller scintillator detector elements as depicted in [Fig. 2a, b]
[9]
[10].
Fig. 2 a, b Schematic drawing of a scintillation detector in top view. Thin septa separate the
detector elements. b While reducing the size of detector elements, the dead space relatively increases
due to the overall larger area of the septa resulting in a decreased dose efficiency
of the detector. c Schematic drawing of a PCD in top view. d & e Side view of EID. The individual detector elements are separated by light-reflecting
septa. f Side view of PCD. The detector pixels are formed by the pixelated anodes without
using additional separating layers. Collimator blades are still necessary to suppress
scatter radiation.
Most PCDs for clinical and preclinical use are composed of semiconductors such as
cadmium telluride (CdTe), cadmium zinc telluride (CdZnTe), and silicon [11]. The semiconductor layer is located between a cathode and a pixelated anode, across
which a high voltage of about 800–1000 V is applied [12]. As shown in [Fig. 1b, ]PCDs convert each incoming X-ray photon directly into electrical signals by generating
a charge cloud of electron-hole pairs. The resulting charge clouds induce current
pulses where the height of the pulses is proportional to the energy of the incoming
X-ray photons. The pulses are then individually counted as soon as they exceed a threshold
and can thus be separated by energy thresholds.
Both, the intensity of the scintillation light and the resulting amplitude of the
induced current pulse are proportional to the energy of the absorbed X-ray photons.
All registered current pulses are integrated over the measurement time for a single
projection. One of the major limitations of this approach is that low-energy X-ray
photons, which carry most of the low contrast information, contribute significantly
less to the integrated detector signal than higher-energy X-ray photons. This energy
weighting substantially reduces the contrast-to-noise ratio (CNR) [13].
The new detector technology is also accompanied by physical challenges, such as k-escape,
Compton scattering, pulse pile-up, and charge sharing, which must be noted. When X-ray
photons arrive in the detector cell, secondary charge clouds can be created in addition
to the primary charge clouds created due to the X-ray fluorescence of the detector
material. When X-ray photons interact with the detector material, Compton scattering
can result in only a portion of the primary energy being deposited as a charge cloud
in the detector element. The remaining energy of the scattered photon can then reach
another detector element. Charge sharing is when a charge cloud is created near a
boundary of two pixel electrodes. As a result, this charge cloud can be measured by
several neighboring pixel electrodes. If two pulses are generated almost simultaneously,
the electrical pulses overlap, and this is called “pulse pile-up”. In this case, incoming
pulses are registered as a single pulse, which in turn leads to an inaccuracy in the
measured photon energy.
Improvement of spatial resolution
Improvement of spatial resolution
The spatial resolution of a CT detector is primarily determined by the size of the
detector elements, which usually range between 0.8 × 0.8 mm2 to 1 × 1 mm2 at the detector level [14]
[15]
[16]
[17]. Increasing spatial resolution of EIDs beyond this point is limited due to the septa
needed to prevent crosstalk between neighboring photodiodes and the reduced quantum
efficiency of the detector, because X-ray photons absorbed in the separating layers
do not contribute to the measured signal ([Fig. 2a, b]) [15].
As shown in [Fig. 2c, ]PCDs, on the other hand, come with smaller pixel sizes as they do not require separating
layers between the detector pixels. Detector elements in PCCT range from 0.11 × 0.11 mm2 to 0.5 × 0.5 mm2. Including a geometric magnification factor, this results in a spatial resolution
of 0.07 × 0.07 mm2 to 0.28 × 0.28 mm2
[14]
[18]
[19]
[20]
[21]. In the standard multi-energy mode, PCD array subpixels are grouped and read out
with the corresponding energy thresholds. In addition, the spatial resolution can
be increased by reading out individual subpixels in the special ultra-high resolution
(UHR) mode.
Increased spatial resolution is of particular importance if subtle changes and small
anatomical structures need to be evaluated such as in chest, bone, and cardiac CT
where clinical benefits have been demonstrated by several preclinical and clinical
investigators already [22]
[23]
[24]
[25]. In [Fig. 3], we provide a representative clinical case of a 73-year-old patient who underwent
cardiac CT to rule out significant coronary artery stenosis.
Fig. 3 Cardiac CT angiography of a 73-year-old patient. a depicts a curved multiplanar reconstruction of the RCA obtained on a second-generation
EID dual-energy CT system, b depicts the same vessel acquired using a PC-CT system, which clearly shows the increased
resolution and improved image quality compared to the EID detector.
[Fig. 4] shows another example of a patient with otosclerosis, which is a slowly progressing
focal disorder of the bone metabolism of the otic capsule. In earlier stages, this
leads to demineralization and spongiotic remodeling. Typical radiologic features comprise
areas of increased bony radiolucency (usually at the fissula ante fenestram), but
also widening of the oval window, thickening of the stapes, and a low-density demineralized
zone outlining the cochlea (the so-called double ring sign) [26]. While the specificity of a high-resolution EID CT is high (around 95 %), its sensitivity
is relatively low with approximately 58 %. In particular, submillimetric, retrofenestral,
and dense sclerotic lesions are difficult to detect on conventional EID CT [27]. However, this can be of particular importance, as it is associated with sensorineural
hearing loss and is treated with a cochlear implant rather than stapedectomy [28].
Fig. 4 Left temporal bone in two patients with fenestral otosclerosis, scanned with PCCT
(NAEOTOM Alpha, Siemens Healthineers, Forchheim, Germany) (a–c, patient A) and conventional high-resolution EID CT (SOMATOM Definition Flash, Siemens
Healthineers, Forchheim, Germany) (d–f, patient B). The bony lucencies adjacent to the oval window are clearly better differentiated
from background noise in (a, coronary) and (b, axial) than in (d, coronary) and (e, axial). In addition, replacement prostheses are more sharply distinguishable in
(c) than in (f).
Image examples of PCCT and EID CT are given in [Fig. 4]. Delineation of the fenestral bony radiolucency is superior on PCCT images. In addition,
the stapes implant ([Fig. 4c]) is more sharply defined on PCCT with fewer artifacts compared to EID CT ([Fig. 4f]). Finally, despite the gain in resolution, the total radiation dose was considerably
lower for PCCT compared to EID CT (CTDIvol 16.6 mGy vs. 34.05 mGy).
Electronic noise
Apart from spatial resolution, image quality is limited by electronic noise in EID
CT. Noise in CT imaging is a composite of quantum noise and electronic noise. The
quantum noise is affected by the number of photons, whereas the electronic noise results
from the electronic circuitry in the system. As mentioned above, an EID measures the
total X-ray energy detected during the time interval of a single projection, which
includes the electronic noise as a random additive term in the measurement. At high
doses, electronic noise is negligible as the quantum noise is proportional to the
incident fluence rate, which in turn increases at high doses [29]
[30]. Conversely, in situations with extremely low radiation doses or in extremely obese
patients, the electronic noise level in EID CT scans will become comparable in strength
to the low detector signal from the X-ray photons, resulting in noise streaks or drift
in Hounsfield Unit stability [29].
Electronic noise usually has a constant low amplitude. Thus, when detected by a PCD,
it can be interpreted as a photon with an energy at the lower end of the typical X-ray
spectrum as shown in [Fig. 5]. This makes it possible to set a threshold and specifically exclude electronic noise
from further signal processing and image reconstruction [31]. However, electronic noise may still have some minor effect on the detected signal
as it artificially increases the energy of the detected photon by increasing its respective
amplitude. Nevertheless, the elimination of electronic noise using a PCD provides
more consistent image quality as shown in [Fig. 3] and noticeably reduces streak artifacts in comparison to EID [32]. Overall, this reduces image noise, improves the diagnostic quality of the acquired
data and improves Hounsfield Unit stability [33]. PCCT may therefore allow new low-dose imaging protocols as electronic noise is
currently the limiting factor with state-of-the-art EID scanners, and might thus be
applied to pediatric imaging or lung cancer screening [22]
[34].
Fig. 5 In a PCD, voltage pulses are induced by the absorbed X-ray photons. These voltage
pulses are counted as soon as they exceed a threshold value (dashed line). The pulse
height corresponds to the energy of the incoming photons by direct conversion. Pulses
with a low amplitude are counted as baseline noise which is caused for example by
electronic noise.
Artifact reduction
Metal artifacts are one of the strongest artifacts encountered in CT. These artifacts
have a characteristic appearance and are caused by various physical processes, such
as photon scattering, photon starvation, and beam hardening. In beam hardening, the
effective photon energy of the X-ray beam, which contains a broad spectrum of energies,
is shifted to the higher end of the spectrum after passing through the scanned object.
These artifacts can have a massive impact on image quality and diagnostic confidence,
and can drastically obscure critical structures of interest. Because PCCT sorts each
photon by its energy, monoenergetic images at high keV levels can be reconstructed
from the multispectral dataset, which allows for a significant reduction in beam hardening
artifacts compared to EID [35]. In addition, the increase in spatial resolution also improves the reduction of
partial volume effects, and thus further reduces artifacts, which might improve the
assessment of small structures with high density such as coronary plaques [24].
The following case of an 85-year-old female patient who received occipito-vertebral
fusion due to an atlas fracture demonstrates this in a typical clinical setting. A
postoperative EID CT examination revealed bilateral resorption margins around the
pedicular screws in C5. However, assessment was impaired due to metal artifact superimposition.
After treatment of delayed wound healing, follow-up CT imaging of the cervical spine
was performed on a PCCT after interim external immobilization of the neck.
Image examples of the two examinations are shown in [Fig. 6]. The clinically established standard is given for EID CT. The PCCT protocol was
designed to mimic the current standard. In addition, inline calculation of virtual
monoenergetic images at 130 keV was performed exploiting the multispectral data to
improve assessment of the metallic implant and surrounding structures. The PCCT images
provide superior image quality and improved assessment of the screw loosening as well
as of adjacent osseous and soft tissue structures at a substantially reduced radiation
dose (CTDIvol 12.45 mGy vs. 7.54 mGy).
Fig. 6 Image example of an 85-year-old patient with posttraumatic cervical spinal fusion.
The loosening of the right pedicle screw in C5 is most obvious in PCCT130keV whereas
artifacts are strongest on EID CT. The radiation dose was substantially lower in the
PCCT examination (CTDIvol 12.45 mGy vs. 7.54 mGy).
Material decomposition
Spectral CT data is acquired by energy separation in PCCT. This data can be reconstructed
in different energy ranges, or the energy information can be used for quantitative
image analysis by energy weighting or material decomposition. In energy weighting,
more weight is assigned to a particular energy bin relative to other energy bins [13]
[33]
[36]
[37]. For material decomposition the full energy dependency of the attenuation curve
in each image voxel needs to be identified [38]
[39]. The underlying hypothesis that any material composed of light elements, such as
human tissue, will have X-ray attenuation properties roughly equivalent to a combination
of two base materials is represented by two bins [39]. Theoretically, any pair of materials can be chosen as base materials. However,
in humans, a combination of water and calcium is assumed. As shown in [Fig. 7a], this allows modelling of each human tissue in a diagram in which the axes represent
the concentrations of the two base materials. Additional dimensions can be added using
the element-specific K-edge for high atomic number elements, where a step-like change
in attenuation occurs at a specific X-ray energy ([Fig. 7b]). This allows the calculation of virtual monochromatic images, but also the distribution
of a certain material in the body, as well as virtual non-contrast images [40] or material-specific color overlay images [40]
[41]
[42].
Fig. 7 Material decomposition. a The X-ray attenuation properties of any material in the human body correspond to
a point in a two-dimensional diagram with water and calcium (Ca2 +) concentrations
on the axes. b Including other K-edges extends the diagram into more dimensions (three shown here).
This makes it possible to measure the contrast agent concentration independently of
the calcium-water concentration.
Furthermore, this approach can be employed for the quantification and separation of
contrast agents. Symons et al. [41] were also able to show that the injection of different contrast agents at different
time points can visualize the incorporation of the contrast agents in only one scan
in the form of a multiphase image. This opens further possibilities, especially with
regard to dose reduction, contrast agent reduction, and tissue characterization. Improving
spectral imaging is also expected to improve the detectability of new contrast agents,
especially k-edge contrast agents [43]. In addition to the currently approved contrast agents with iodine and gadolinium,
further contrast agents are required to generate multiphase images, which can also
be based on nanoparticles. However, more research is needed to pave the way to clinical
practice and to investigate the true benefits with respect to improving patient care.
The amount of contrast agent in an image voxel can be separated from the other components
if three parameters are measured: the concentrations of water, calcium, and the contrast
agent. To determine these three variables, measurements in three or more energy ranges
are required [38] as shown in [Fig. 7b]. Existing methods for measuring iodine concentration at two energies must therefore
rely on a priori assumptions about tissue composition. PCCT introduces the potential
for more accurate measurements [6]. For example, iodine exhibits the K-edge at 33 keV. In order to perform independent
quantification of iodine, it is necessary for photons to be transmitted at these low
energies. This seems realistic for objects or patients with small diameters such as
children [41]
[42]
[44]
[45].
Data storage and postprocessing
Data storage and postprocessing
Due to the improved spatial resolution and new possibilities for quantitative image
postprocessing, PCCT will substantially increase the amount of acquired data. Especially
when working with larger matrices like 1024 × 1024 in combination with multi-spectral
photon-counting CT data, the amount of data increases significantly. For example,
reducing the slice thickness by 50 % and doubling the in-plane matrix increases data
size by eight-fold per reconstruction method. As a result, datasets of a single patient/examination
are likely to regularly exceed 10 gigabytes especially when exploiting the entire
multispectral data for advanced postprocessing, such as calculating virtual monochromatic,
virtual non-contrast, or virtual non-calcium images or performing different material
decompositions. Therefore, new solutions for data handling, transfer, storage, and
presentation as well as new algorithms for processing and analyzing data are required.
With recent advances in artificial intelligence, new approaches for automated high-throughput
data management and analysis will become available. Implementing such methods into
clinical workflows has the potential to support and accelerate the clinical potential
of PCCT in daily practice. Besides commercial solutions provided by all major vendors,
several open-source tools like 3 D Slicer [46] or pyradiomics [47] are available for image analysis and are established in the research community.
More recently developed solutions, such as JIP (DKFZ German cancer consortium, Heidelberg,
Germany) and NORA (NORA Medical Imaging Platform Project, University Medical Center
Freiburg, Department of Radiology, Freiburg, Germany), enable a platform approach,
which makes data annotation and postprocessing solutions more accessible, modular,
and user-friendly, even across multiple institutions. This has the potential to improve
today’s workflows, refine clinical-decision making and personalize patient management.
As of now, these solutions have been investigated in various research settings to
explore their potential role to improve clinical workflows and patient management.
The most promising results were found for fully automated organ and tissue segmentation,
extraction of quantitative radiomic imaging features, which may facilitate improved
tissue characterization and end-to-end deep learning pipelines for individualized
risk assessment. However, further research is needed to prove their value in clinical
scenarios.
Moreover, since the implementation of more complex and iterative image reconstruction
algorithms, the use of traditional image quality metrics such as SNR and CNR for objective
image assessment in PCCT remains limited due to the nonlinearity of iterative reconstruction
methods [48]. This is also present for different model-based and deep learning reconstructions
[49] offered by manufacturers. Further development of new, robust methods will be necessary
in the future.
Conclusion
PCCT has been implemented in the clinical routine and the novel detector technology
significantly decreases image noise and artifacts, improves spatial resolution, and
reduces radiation dose. In addition, K-edge imaging with material decomposition creates
new possibilities for quantitative analyses. To exploit the full potential of PCCT,
reliable and automated tools are required to support data analyses and establish efficient
and accurate ways for data postprocessing, handling, and storage.
