{
  "id": "photon-counting-computed-tomography",
  "title": "Photon-counting Computed Tomography",
  "lang": "en",
  "source_lang": "en",
  "summary": "Photon-counting computed tomography uses energy-resolving X-ray detectors to produce detailed cross-sectional images and distinguish materials by their spectral properties.",
  "category": "health",
  "subcategory": "medicine",
  "type": "concept",
  "aliases": [
    "Photon-counting CT",
    "PCCT",
    "PCD-CT",
    "Photon-counting detector CT",
    "Spectral photon-counting CT"
  ],
  "generated_at": "2026-11-04",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/photon-counting-computed-tomography/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/photon-counting-computed-tomography/",
      "json": "https://aiwiki.page/zh/photon-counting-computed-tomography.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "Photon-counting computed tomography is a form of X-ray computed tomography in which detectors register individual incoming X-ray photons and classify them according to estimated energy. Conventional clinical CT detectors measure the combined energy deposited by many photons during a sampling interval. Photon-counting detectors instead generate and process separate electrical pulses, preserving information about both the number and the energy distribution of detected photons.\n\nThis change in detector operation supports smaller detector elements, energy-selective imaging, and suppression of electronic noise below a chosen detection threshold. Its principal clinical applications combine high-resolution anatomical imaging with spectral measurements that help distinguish substances such as iodine, calcium, and water. These capabilities are relevant to cardiovascular, pulmonary, abdominal, musculoskeletal, and neurological imaging, although their practical benefits depend on the scanner, acquisition protocol, reconstruction method, and diagnostic task.\n\nPhoton-counting CT retains the rotating X-ray source, projection acquisition, and tomographic reconstruction principles of conventional CT. Its distinguishing features lie in detector physics and the information extracted from each exposure. Commercial clinical deployment began in 2021 after decades of development in semiconductor detectors, high-speed electronics, and spectral reconstruction."
    },
    {
      "heading": "Physical principles and signal formation",
      "text": "A CT examination measures how an X-ray beam is attenuated as it passes through the body from many directions. Attenuation depends on photon energy and the composition and thickness of the traversed tissues. In the diagnostic energy range, the principal interactions are the photoelectric effect and Compton scattering. Their different energy dependencies provide a physical basis for distinguishing materials, rather than representing every tissue only by a single CT number.\n\nIn an idealized energy-resolving detector, the expected number of photons recorded in an energy interval can be written as\n\n\\[N_k=\\int_{E_k}^{E_{k+1}}\\Phi_0(E) \\exp\\!\\left[-\\int_L\\mu(E,\\mathbf r)\\,dl\\right]\\eta(E)\\,dE,\\]\n\nwhere \\(\\Phi_0(E)\\) is the incident photon spectrum, \\(\\mu(E,\\mathbf r)\\) is the energy-dependent linear attenuation coefficient, \\(L\\) is the ray path, and \\(\\eta(E)\\) is the detection efficiency. The interval boundaries \\(E_k\\) and \\(E_{k+1}\\) define an energy bin. Real detectors require a more elaborate response model because the measured pulse height does not always correspond exactly to the incident photon energy.\n\nConventional energy-integrating detectors usually contain a scintillator that converts X-rays into visible light, followed by a photodiode that converts the light into an electrical signal. Their output is approximately weighted by the energy deposited: a higher-energy photon contributes more signal than a lower-energy photon. Photon-counting detectors generally use a semiconductor that converts deposited X-ray energy directly into mobile electrical charge. An applied electric field moves the charge carriers toward electrodes, producing a pulse whose amplitude is related to deposited energy.\n\nA discriminator compares each pulse against one or more energy thresholds. Pulses exceeding the lowest threshold are counted, while additional thresholds divide the detected spectrum into categories. In an ideal counting measurement, each accepted photon contributes one count, rather than a contribution proportional to its energy. This can improve the relative contribution of lower-energy photons, which often carry strong contrast information for iodine and soft tissues. It does not mean that low-energy photons are always useful: many are absorbed before reaching the detector, especially in larger patients.\n\nThe detector measures deposited energy, not photon energy with perfect precision. Charge transport, electronic fluctuations, fluorescence escape, and interactions involving neighboring pixels broaden or distort the measured spectrum. Consequently, photon-counting CT is an energy-discriminating technology rather than an ideal spectrometer. Its useful spectral performance depends on how accurately the complete imaging chain models these departures from ideal counting."
    },
    {
      "heading": "Detector materials, pixels, and scanner architecture",
      "text": "Clinical photon-counting detectors must combine efficient X-ray absorption with rapid charge collection and stable operation under intense irradiation. Cadmium telluride and cadmium zinc telluride are important detector materials because their relatively high atomic numbers and densities allow efficient absorption within a compact detector thickness. Silicon is also used in photon-counting detector development. Its lower stopping power can be compensated by an edge-on geometry, in which photons travel along a comparatively long path through the sensor while charge is collected across a much shorter distance.\n\nThe absorber is coupled to an application-specific integrated circuit that amplifies pulses, compares them with thresholds, and accumulates counts. These electronics must operate within the small area assigned to each detector pixel. Practical designs balance pixel dimensions, power consumption, heat removal, energy discrimination, and maximum count rate. The number of thresholds and available acquisition modes are therefore engineering choices rather than fixed attributes of photon-counting CT.\n\nPixel size is especially important. Energy-integrating detectors commonly require reflective separators between scintillator elements to limit optical cross-talk. Direct-conversion detectors do not need these particular separators, which facilitates smaller detector elements without the same loss of active area. Nevertheless, photon-counting detector arrays still contain electrodes, interconnections, support structures, and sometimes anti-scatter components. Their geometric efficiency is high, but not unlimited.\n\nSmaller pixels reduce the number of incident photons per pixel per unit time, helping electronics distinguish successive events. They also improve spatial sampling. However, as pixels become smaller, a greater fraction of interactions occurs close to a boundary, increasing the likelihood that charge from one photon will be shared among multiple pixels. Pixel miniaturization therefore improves some aspects of performance while making others more difficult.\n\nDetector-element dimensions must be distinguished from reconstructed image resolution. Dimensions quoted at the detector differ from dimensions projected to the scanner’s isocenter because of geometric magnification. Final resolution also depends on focal-spot size, acquisition geometry, patient motion, reconstruction kernel, and slice thickness. Some clinical photon-counting systems provide nominal section thicknesses around 0.2 mm in dedicated high-resolution modes, but this is not a universal specification or a guarantee that every structure of that size will be resolved.\n\nPhoton-counting CT can use single-source or dual-source scanner architectures. In a dual-source design, two X-ray tube–detector systems provide complementary angular information and can improve temporal resolution. Temporal resolution is not an intrinsic consequence of photon counting: it depends principally on rotation speed, source arrangement, and reconstruction strategy. Likewise, scan coverage and acquisition speed depend on the complete scanner design, not solely on detector technology."
    },
    {
      "heading": "Count-rate limitations and spectral imperfections",
      "text": "Diagnostic CT exposes detectors to much higher photon fluxes than many laboratory spectroscopy applications. Each pixel may need to process millions of events per second, with substantial variation across detector position, projection angle, and patient anatomy. If two photons arrive within the electronics’ resolving time, their pulses may overlap. This phenomenon, called pulse pile-up, is a central constraint on clinical photon counting.\n\nPile-up can cause both counting losses and spectral distortion. Two nearly simultaneous photons may produce a single pulse with an amplitude resembling the sum of their deposited energies. In other circumstances, pulse overlap or detector dead time prevents one or more events from being registered. The resulting errors depend on the detector response and incident spectrum, so a simple global correction is generally inadequate. High-flux regions and weakly attenuated rays can be particularly demanding even when the patient’s overall exposure is not unusually high.\n\nCharge sharing produces a different class of error. An X-ray interaction creates a cloud of charge carriers that can spread across neighboring electrodes. Without correction, a single photon may be counted as two lower-energy events, or part of its charge may fall below a threshold and go unrecorded. Electronic charge-summing methods can combine coincident signals in adjacent pixels, but they require additional circuitry and can themselves become difficult to operate at high count rates.\n\nCharacteristic fluorescence also redistributes energy. Following photoelectric absorption, a detector atom may emit a secondary X-ray that escapes the original interaction site. If that X-ray leaves the sensor, the measured energy is lower than the incident photon energy. If it is absorbed in another pixel, it can create an additional event elsewhere. These processes are especially relevant in high-atomic-number detector materials and contribute to correlations among pixels and energy bins.\n\nMaterial defects, charge trapping, incomplete charge collection, and pixel-to-pixel threshold differences add further departures from ideal behavior. Some semiconductor sensors can also experience changes in internal electric field under sustained irradiation, often described as polarization. Detector design, operating conditions, material quality, and correction procedures all influence the severity of these effects.\n\nCorrection therefore combines hardware design with calibration and mathematical modeling. A detector response function can describe the probability that a photon of a given energy will be recorded in a particular measured-energy bin, potentially including redistribution to neighboring pixels. Reconstruction and material decomposition can incorporate this response. Residual errors matter because an image may remain visually acceptable while its spectral measurements are biased; anatomical image quality and quantitative material accuracy are related but distinct performance criteria."
    },
    {
      "heading": "Spectral imaging and material decomposition",
      "text": "Spectral CT uses the energy dependence of attenuation to obtain information beyond a conventional grayscale image. Photon-counting detectors acquire energy-selective data during the same exposure, although the number of bins actually retained and reconstructed depends on the scanner and acquisition mode. Unlike approaches that alternate tube voltage or combine two differently filtered X-ray spectra, photon counting sorts transmitted photons at the detector.\n\nA common model expresses attenuation as a combination of basis functions:\n\n\\[\\mu(E,\\mathbf r)\\approx \\sum_{j=1}^{m}a_j(\\mathbf r)f_j(E).\\]\n\nThe functions \\(f_j(E)\\) represent selected material attenuation curves or physical interaction components, while the coefficients \\(a_j(\\mathbf r)\\) describe their local contributions. A two-material model might use water and iodine as its basis. The coefficients are estimated from energy-dependent projection measurements, either before image reconstruction or through a joint reconstruction process. Image-domain decomposition is another approach, though its noise behavior and correction requirements differ.\n\nOne important output is the virtual monoenergetic image, which estimates the appearance of the object at a selected photon energy. Low-energy images can increase iodine contrast because iodine attenuation rises strongly toward its K absorption edge, approximately 33.2 keV. Higher-energy images can reduce some forms of beam-hardening and metal-related streaking, although neither improvement is universal. A virtual monoenergetic image remains a model-derived reconstruction; the patient was not actually scanned with a monochromatic beam.\n\nMaterial maps estimate the distribution of a selected substance, commonly an iodine-based contrast agent. Iodine maps can show enhancement that is less apparent on conventional images and provide quantitative concentration estimates under suitable conditions. Virtual noncontrast images attempt to remove iodine mathematically from contrast-enhanced data. Their accuracy varies with anatomy, contrast concentration, calcification, lesion composition, and reconstruction method, so they are not automatically interchangeable with a true unenhanced acquisition.\n\nA further possibility is K-edge imaging, which exploits the abrupt rise in attenuation above an element’s inner-shell binding energy. Gadolinium has a K edge near 50.2 keV and gold near 80.7 keV. With sufficiently informative measurements, such features can support identification of a contrast material or separation of multiple agents. Much of this work remains investigational because useful spectral separation must be combined with adequate photon counts, acceptable biological properties, and clinically feasible contrast administration.\n\nMore thresholds do not automatically produce more reliable material information. Dividing a fixed photon population into additional bins reduces counts per bin, while detector spectral blurring makes neighboring bins less independent than their nominal boundaries suggest. Threshold selection and material models must therefore match the intended task. The number of materials that can be estimated reliably is determined by usable independent spectral information, noise, and modeling assumptions, not simply by the number of displayed image series."
    },
    {
      "heading": "Image reconstruction, resolution, and quantitative performance",
      "text": "Photon-counting CT reconstruction begins with energy-bin measurements collected over many projection angles. Processing includes detector corrections, normalization, scatter and spectral corrections, and conversion into quantities suitable for reconstruction. Conventional attenuation images can be reconstructed from suitably combined counts, while spectral products require additional energy-dependent modeling. Filtered back projection, iterative reconstruction, and learned reconstruction methods may all be used within this framework.\n\nIn ideal counting, the variance of a measured count is approximately equal to its mean, following a Poisson distribution. At low counts, relative statistical fluctuations become large. Actual detector data can depart from independent Poisson behavior because of pile-up, charge sharing, cross-talk, and correction procedures. Reconstruction methods that account for the measurement statistics can reduce noise, but the assumptions must reflect the detector’s real behavior.\n\nSpatial resolution is often characterized with the modulation transfer function, which describes how contrast is transferred at different spatial frequencies. Noise is characterized not only by its standard deviation but also by its texture and spatial-frequency distribution. A sharper reconstruction kernel may improve the visibility of small boundaries while increasing image noise. Conversely, strong smoothing can produce a visually quiet image while obscuring fine structures. The relevant comparison is therefore performance for a defined detection or measurement task.\n\nSmall detector elements can reduce partial-volume effects, in which several tissues contribute to the same reconstructed voxel. This is particularly important at the interfaces of small vessels, calcifications, stents, and thin bone. Reduced partial-volume averaging can limit apparent enlargement, or blooming, of dense structures. However, patient motion, focal-spot blur, and beam-hardening artifacts can still dominate. Detector resolution alone does not remove these other sources of error.\n\nIterative methods incorporate a forward model of image formation and a penalty or prior that constrains the reconstructed image. Artificial intelligence methods can also be used for denoising or reconstruction, including models trained to map noisy data to reference-quality images. Their performance depends on training data, reconstruction settings, and the resemblance between training conditions and clinical cases. Noise reduction is not by itself evidence that every subtle lesion or texture feature has been preserved.\n\nQuantitative assessment extends beyond conventional Hounsfield units to iodine concentration, material fractions, virtual monoenergetic attenuation, and other derived measurements. Reproducibility requires attention to patient size, spectrum, reconstruction kernel, dose, and contrast timing. A numerical material map may be affected by motion or model mismatch even when the corresponding anatomical image looks plausible. Standardized phantoms and cross-platform comparisons are consequently important for separating physical measurement accuracy from favorable image appearance."
    },
    {
      "heading": "Radiation exposure and contrast efficiency",
      "text": "Photon-counting CT has several mechanisms that can improve the use of detected radiation. A lower energy threshold can exclude electronic noise pulses that do not correspond to accepted X-ray events. Direct conversion avoids the intermediate scintillation-light stage, while energy-selective processing can place greater emphasis on photons that carry useful information for a particular task. These features can improve dose efficiency, especially when conventional detector electronic noise would otherwise contribute substantially.\n\nThe improvement is not a fixed percentage applicable to every examination. Radiation requirements depend on patient size, anatomical region, diagnostic task, detector mode, reconstruction, and the image quality accepted for interpretation. A protocol designed to reveal smaller structures may use the detector’s performance to increase resolution rather than reduce exposure. In other circumstances, a similar diagnostic task may be achievable at a lower scanner output. These are different uses of the same technical capability.\n\nReported comparisons therefore require careful matching. A low-dose image reconstructed with a strong denoising method cannot be interpreted solely from its noise level without considering spatial resolution and lesion visibility. Likewise, comparing a new photon-counting protocol with an older conventional protocol may mix detector effects with changes in tube voltage, reconstruction, or scan coverage. Task-based phantom studies and appropriately designed clinical comparisons help identify which component produces an observed advantage.\n\nCT exposure is commonly described using the volume CT dose index, \\(CTDI_{\\mathrm{vol}}\\), and dose–length product, or DLP. These are standardized scanner-output quantities, not direct measurements of an individual patient’s organ doses. Size-specific dose estimates adjust an output-based estimate for patient dimensions. Radiation dosimetry remains necessary because photon-counting detectors do not change the fact that CT uses ionizing radiation.\n\nSpectral processing may also improve iodine efficiency. Lower-energy virtual monoenergetic images can preserve conspicuous vascular or tissue enhancement with less iodine under some conditions. However, iodine concentration, total iodine delivery, injection rate, and acquisition timing interact with cardiac output and anatomy. Stronger apparent enhancement is not equivalent to a demonstrated improvement in diagnostic accuracy, nor does it establish a universal reduction in contrast-related clinical risks.\n\nVirtual noncontrast reconstruction can sometimes eliminate a separate unenhanced acquisition, creating an additional route to exposure reduction. Whether this is valid depends on the diagnostic question. Small calcifications, hemorrhage, fat-containing lesions, and highly concentrated iodine can challenge material separation. The relevant safety framework is patient safety across the complete examination: sufficient information must be retained to avoid nondiagnostic scans, missed findings, or unnecessary repeat imaging."
    },
    {
      "heading": "Cardiovascular and thoracic applications",
      "text": "Cardiovascular imaging combines several demanding requirements: high spatial resolution for small vessels, high temporal resolution for cardiac motion, and sufficient contrast between blood, vessel wall, plaque, and surrounding tissue. Photon-counting CT addresses the spatial and spectral components directly. When combined with a fast scanner architecture and electrocardiographic synchronization, it can provide detailed coronary images, but the detector alone does not freeze cardiac motion.\n\nIn coronary artery disease, calcified plaque can appear larger than its true dimensions because of partial-volume averaging and limited spatial resolution. This blooming may obscure the residual vessel lumen and complicate estimation of narrowing. High-resolution photon-counting acquisitions can reduce blooming, making the lumen more visible in selected cases. Severe calcification, motion, and image noise nevertheless remain potential limitations, and improved visibility does not make every stenosis measurement accurate.\n\nCoronary stents present a related problem. Their dense, thin struts can obscure the lumen through blooming and other artifacts. Smaller reconstructed voxels and sharper kernels can improve visualization, with results depending on stent diameter, strut material and thickness, motion, and scan conditions. Imaging of small stents remains more difficult than imaging of larger ones. Spectral and high-resolution outputs can provide complementary information, but the best settings for fine structural detail are not necessarily those that minimize noise.\n\nIn pulmonary imaging, thin sections can depict small airways, fine reticulation, interlobular septa, and the walls of cystic spaces. Such structures are relevant to fibrotic lung disease, bronchiectasis, and emphysema assessment. Improved spatial sampling may also refine measurements of pulmonary nodules. These uses require control of inspiration, motion, reconstruction texture, and measurement variability, particularly when images are compared across serial examinations.\n\nContrast-enhanced thoracic applications include assessment of pulmonary embolism and aortic disease. Iodine maps can display the regional distribution of contrast in the lungs, providing information associated with blood delivery. They are not direct measurements of perfusion: enhancement at one acquisition time is influenced by contrast timing, motion, gravity, cardiac function, and artifacts. Their interpretation remains linked to the vascular and anatomical images.\n\nEvidence in cardiovascular and thoracic imaging includes technical measurements, reader assessments, and diagnostic comparisons. These address different questions. Better visibility of plaque, stent lumens, or lung microstructure does not by itself demonstrate fewer cardiovascular events or improved survival. Establishing such downstream benefits requires study designs and follow-up beyond a comparison of image sharpness or reader confidence."
    },
    {
      "heading": "Abdominal, musculoskeletal, and neurological applications",
      "text": "Abdominal photon-counting CT combines soft-tissue contrast, spectral characterization, and potentially reduced partial-volume effects. Low-energy virtual monoenergetic images can increase the conspicuity of iodine-enhanced vessels and lesions. Iodine maps provide a separate representation of enhancement, which can help distinguish enhancing tissue from intrinsically dense material. Their interpretation depends on acquisition phase, because arterial, portal venous, and delayed images represent different physiological conditions.\n\nApplications in cancer imaging include detection and characterization of lesions in the liver, pancreas, and other organs. A small lesion may become more conspicuous if its enhancement differs from the surrounding tissue, but spatial resolution cannot compensate for absent contrast or unsuitable timing. Quantitative iodine concentration may serve as an imaging biomarker, although validation requires reproducibility and evidence connecting the measurement to a defined biological or clinical outcome.\n\nUrinary stone analysis is a well-established spectral CT task to which photon counting can contribute. Kidney stones of different chemical composition exhibit different energy-dependent attenuation. Separation of uric-acid from non-uric-acid stones is more established than reliable identification of every possible stone constituent. Small stones, mixed composition, patient size, and partial-volume effects can affect classification. More detailed energy sampling does not eliminate these sources of uncertainty.\n\nMusculoskeletal applications exploit the high contrast of bone and the fine dimensions of many clinically relevant structures. High-resolution imaging can improve depiction of cortical surfaces, trabeculae, subtle fractures, and small joints. In the temporal bone, structures such as the ossicles and the bony boundaries of the inner ear are especially demanding. These examinations can benefit from fine spatial sampling, although imaging of low-contrast soft tissues remains a different task from resolving high-contrast bone.\n\nSpectral methods can help reduce some metal-related artifacts around orthopedic hardware and support material-selective reconstructions. Higher-energy virtual monoenergetic images may reduce streaking, but they can also reduce soft-tissue contrast. Virtual noncalcium techniques attempt to suppress mineralized bone and reveal differences in the underlying marrow, including changes associated with edema. Their performance depends on the decomposition model and does not make them equivalent to magnetic resonance imaging.\n\nNeurological uses include head and neck angiography, evaluation of small vessels, and distinction between iodine and blood after selected vascular procedures. Spectral separation is relevant when contrast staining and intracerebral hemorrhage have overlapping appearances on conventional CT. The skull, beam hardening, motion, and low soft-tissue contrast remain important constraints. Photon counting expands available measurements, but the brain’s subtle tissue differences still demand careful validation of reconstruction and diagnostic performance."
    },
    {
      "heading": "Development, clinical evidence, and implementation",
      "text": "Photon-counting CT emerged from several intersecting developments: energy-resolving radiation detectors, semiconductor fabrication, fast integrated electronics, and methods for spectral reconstruction. Material-selective CT itself predates clinical photon-counting scanners. In 1976, Robert Alvarez and Albert Macovski described an energy-selective reconstruction framework that helped establish the theoretical basis for representing attenuation using a small number of energy-dependent components. Later detector advances made it increasingly practical to acquire such information at clinical CT count rates.\n\nEarly photon-counting systems were constrained by limited count-rate capability, detector area, stability, and data-handling capacity. Small-animal systems, laboratory scanners, and specialized geometries provided opportunities to test spectral and high-resolution imaging before full clinical implementation. Whole-body research systems subsequently enabled human studies and comparisons with established CT technology. Translation required not just successful photon counting, but reliable operation across the full range of patient sizes and scan protocols.\n\nIn September 2021, the United States Food and Drug Administration cleared Siemens Healthineers’ NAEOTOM Alpha, the first commercially available photon-counting CT system. This marked a transition from research prototypes to routine clinical procurement and use. Regulatory clearance established a commercial milestone; it did not establish superiority for every clinical indication. Hardware designs, available modes, and reconstruction software have continued to develop.\n\nClinical evidence must distinguish among physical performance, diagnostic performance, management effects, and patient outcomes. A phantom study can establish spatial resolution or material-measurement accuracy under controlled conditions. Reader studies assess visibility or confidence. Studies of diagnostic sensitivity and specificity require suitable reference standards, while changes in treatment decisions or outcomes require additional follow-up and appropriate comparison groups. Results from one protocol or organ system cannot automatically be transferred to another.\n\nImplementation introduces substantial workflow requirements. A single examination can produce conventional images, multiple monoenergetic series, material maps, and high-resolution reconstructions. These increase reconstruction workload, storage needs, and the volume of images available for interpretation. Protocol design determines which products are generated routinely and which are produced only when needed. Radiologists and technologists must also recognize artifacts specific to energy-bin processing and material decomposition.\n\nOutstanding questions concern reproducibility across platforms, performance in large patients, quantitative standardization, and the value of additional spectral information for specific decisions. Cost-effectiveness analysis must account for acquisition and maintenance costs, examination throughput, potential replacement of other tests, and any demonstrated reduction in repeat imaging. Emerging multi-agent and K-edge applications additionally depend on contrast-agent development and safety evaluation, not just detector capability. Their clinical adoption therefore requires coordinated advances in imaging physics, clinical trials, and regulated pharmaceutical development."
    }
  ],
  "links": [
    "x-ray-computed-tomography",
    "photon",
    "photoelectric-effect",
    "compton-scattering",
    "semiconductor",
    "calibration",
    "contrast-agent",
    "poisson-distribution",
    "artificial-intelligence",
    "patient-safety",
    "coronary-artery-disease",
    "fibrosis",
    "cancer",
    "pancreas",
    "biomarker",
    "kidney-stones",
    "edema",
    "magnetic-resonance-imaging",
    "intracerebral-hemorrhage",
    "diagnostic-sensitivity-and-specificity",
    "clinical-trial"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}