X-ray Radiography

X-ray radiography, also called projection radiography or plain-film X-ray, makes a two-dimensional image of the inside of the body by sending X-rays through it and recording, at each point of a detector, how many arrive. The image records the “shadows” of the structures along each path (NIBIB, Medical X-rays): every pixel sums the attenuation along one ray, so structures at different depths are superimposed. Computed Tomography uses the same physics but measures many projections and computes cross-sections from them; radiography stops at the single projection. That makes it quick, a single exposure, and it makes the image hard to read, for people and for models alike, because depth is lost and overlapping structures add up.

For computer vision the radiograph is a deceptively familiar input. It looks like a greyscale photograph, but its geometry is a central projection from a point source with magnification that depends on depth, its intensities are logarithms of photon counts shaped by the beam’s spectrum and by scattered radiation, and the image a model receives has usually been through vendor-specific processing designed for human readers. This article covers the sensor from the tube to the stored DICOM object. Chest radiograph and mammography analysis, the main applications, are subjects in their own right; Computer Vision in Radiology covers how models are built and evaluated on these images.

How It Works

Producing X-rays

An X-ray tube is an evacuated tube with a heated cathode and a metal anode. Electrons leave the hot filament, are accelerated by the tube voltage and are slowed down in the anode, and the deceleration produces X-rays (Berger et al., 2018). Two processes contribute. Bremsstrahlung, from electrons deflected by the field of a nucleus, gives a continuous spectrum, because an electron can lose any amount of its energy up to all of it. Characteristic radiation, produced when an incoming electron removes an inner-shell electron and an outer-shell electron fills the vacancy, gives lines at energies fixed by the anode material (Berger et al., 2018).

The tube voltage sets the top of the spectrum: each kilovolt gives each electron 1 keV, and no photon can carry more energy than the electron that made it, so the maximum photon energy in keV equals the peak voltage in kV, written kVp (Sprawls, X-ray Production). Sprawls describes the unfiltered bremsstrahlung spectrum as roughly a triangle, with the number of photons rising as energy falls. Most tubes use tungsten anodes; mammography tubes use molybdenum, whose characteristic lines at 17.9 and 19.5 keV fall where breast imaging needs them, and some add rhodium, whose slightly higher-energy radiation is more penetrating for dense breasts (Sprawls, X-ray Production). A thin metal plate, the X-ray filter (not to be confused with an image filter), removes most low-energy photons before they reach the patient, since the body would absorb almost all of them, adding dose without improving the image (Berger et al., 2018).

The electrons strike a small area of the anode, the focal spot, usually roughly rectangular and 0.1–2 mm across. Most tubes offer a small and a large focal spot: the small one blurs less, the large one tolerates more heat, and mammography uses small ones (Sprawls, X-ray Production).

Attenuation and the Beer–Lambert law

At diagnostic energies photons interact with tissue by the photoelectric effect, which absorbs the photon completely, by Compton scattering, which deflects it with a partial loss of energy and dominates in tissue-like materials at higher energies, and by Rayleigh scattering, which changes its direction without loss and contributes little at imaging energies (Berger et al., 2018). Photons that pass without interacting are the primary beam, and they form the image.

For a monochromatic beam crossing a material with linear attenuation coefficient $\mu$ along a path of length $x$, Berger et al. write Lambert–Beer’s law as $I = I_0 e^{-\mu x}$, and for a ray through an inhomogeneous object

$$-\ln \frac{I}{I_0} = \int \mu(x)\, dx ,$$

with $I_0$ the unattenuated intensity (Berger et al., 2018, Eqs. 7.3–7.4). Computed Tomography develops this line integral into the Radon transform and reconstruction; a radiograph simply displays it, one ray per pixel. A real tube emits a spectrum, and $\mu$ depends on energy, so the detected intensity is a sum over energies,

$$I = \int_0^{E_\text{max}} I_0(E)\, \exp\left(-\int \mu(x, E)\, dx\right) dE ,$$

where Berger et al. normalize the spectrum so that $\int I_0(E)\, dE = 1$ (Eq. 7.5). Taubmann et al. (2018, Eq. 8.48) add a detector sensitivity $D(E)$ inside the integral, since a detector does not weight all energies equally. Because low-energy photons are absorbed first, the transmitted spectrum shifts towards higher energies as the path grows, so the logarithm is no longer proportional to the path length (Taubmann et al., 2018). This is beam hardening. In a radiograph it has no streaks to show, as it does in CT, but it means that the same bone produces a different logarithmic contrast depending on how much tissue lies in front of and behind it, which the example below measures.

Scatter and anti-scatter grids

Compton-scattered photons leave the patient in new directions and land on the detector away from the line through their point of origin. They add a broad, slowly varying haze that carries little information about where they were scattered. For a small detail that changes the primary signal $P$ by a fraction $c$, a uniform scatter signal $S$ reduces the measured fractional contrast to

$$\frac{c\,P}{P + S} = \frac{c}{1 + \text{SPR}}, \qquad \text{SPR} = \frac{S}{P},$$

where SPR is the scatter-to-primary ratio. This follows directly from adding $S$ to both the detail and its surroundings. The scattered photons also bring Poisson noise of their own, so at a fixed primary dose the contrast-to-noise ratio falls by $\sqrt{1 + \text{SPR}}$; both effects appear in the example.

An anti-scatter grid reduces scatter: thin lead strips separated by a material transparent to X-rays, placed in front of the detector and aligned towards the focal spot, so that unscattered photons pass between the strips while most scattered photons, arriving at oblique angles, are absorbed (Taubmann et al., 2018). The grid costs dose, because the exposure must be raised to compensate (the Bucky factor, which the report calls a dose penalty), and a grid tilted relative to the beam absorbs primary radiation as well, called grid cutoff, which AAPM Report No. 93 singles out as a problem in bedside radiography (AAPM, 2006). The report also notes that there is no universal agreement on grid type, ratio, interspace material or strip frequency for digital detectors, and that higher-frequency grids need higher ratios for the same scatter cleanup, at a higher dose penalty.

Detectors

The detector turns the transmitted X-rays into an image, and its generations explain much of the variety in radiograph datasets.

• Screen-film. Film is far more sensitive to light than to X-rays, so film is sandwiched between intensifying screens that absorb X-rays and convert about 5–20% of the absorbed energy into light; a thicker screen absorbs more but blurs more, and single-screen, single-emulsion film is used where detail matters most, as in mammography (Sprawls, Screen/Film Radiographic Receptors). The film is both the detector and the display, with a narrow useful range of exposure (AAPM, 2009).
• Computed radiography (CR), or photostimulable storage phosphor imaging. An imaging plate in a film-like cassette stores absorbed X-ray energy as trapped electrons, a latent image proportional to the incident photons. A reader scans the plate with a laser, the released light is collected by a photomultiplier and digitized, and the plate is erased with intense light and reused (AAPM, 2006). The plate responds proportionally over more than four orders of magnitude of exposure, and the signal is amplified logarithmically or by a square root, either before digitization or digitally after it (AAPM, 2006). DICOM gives “digital luminescence radiography” as an equivalent term (NEMA, 2026, A.2).
• Flat-panel detectors (DR). Berger et al. (2018) date flat panels to the mid-1990s and describe them as the state of the art for radiography and angiography. Indirect-conversion panels convert X-rays to light in a layer of caesium iodide, grown as columns that guide the light like optical fibres, and read it with a matrix of photodiodes; direct-conversion panels absorb X-rays in a photoconductor and collect the released charge with an electric field. Each pixel has a thin-film transistor, and the panel is read out row by row. Berger et al. give side lengths of up to 40 cm, pixels of about 100–150 µm and 12- to 16-bit digitization for common medical panels.
• Image intensifiers, used in fluoroscopy before flat panels, convert X-rays to light, light to electrons, accelerate and focus the electrons onto an output phosphor, and view that with a camera (Berger et al., 2018).

Digital detectors have largely replaced screen-film systems (NIBIB, Medical X-rays).

Image geometry

A radiograph is a central projection from the focal spot onto the detector, the same geometry as the pinhole camera model with the source as the centre of projection, except that the object lies between the centre and the image plane. With source-to-image distance SID and source-to-object distance SOD, an object is magnified by

$$M = \frac{\text{SID}}{\text{SOD}},$$

which is how DICOM defines the Estimated Radiographic Magnification Factor (NEMA, 2026, C.8.11.5). Structures farther from the detector are magnified more, so a single radiograph has no single scale. A focal spot of width $f$ is not a point, and similar triangles give a blur at the detector, the geometric unsharpness, of

$$U = f\,\frac{\text{SID} - \text{SOD}}{\text{SOD}} = f\,(M - 1),$$

so blur grows with magnification and vanishes for an object on the detector. This is our derivation from the geometry; Sprawls (X-ray Production) states the qualitative result that small focal spots blur less. Focal-spot, receptor and motion blur combine, and when one is much larger than the others the total is essentially that largest blur (Sprawls, Blur, Resolution, and Visibility of Detail).

Data It Produces

A radiograph is one 2D array of unsigned integers per exposure. DICOM’s Digital X-Ray Image IOD requires that an image be the result of a single exposure, so that its orientation attributes stay meaningful, and deprecates the screen-film and computed radiography practice of several exposures on one cassette or plate (NEMA, 2026, A.26). Its DX Image Module allows 6 to 16 stored bits per pixel and two photometric interpretations (NEMA, 2026, C.8.11.3): MONOCHROME2, where the minimum value is displayed as black after the grey-scale (VOI) transformation, and MONOCHROME1, where it is displayed as white (NEMA, 2026, C.7.6.3.1.2).

Several attributes decide how pixel values should be read:

• For processing or for presentation. The DX IOD is used in two storage classes, distinguished by Presentation Intent Type: images intended for presentation, and images intended for further processing first (NEMA, 2026, A.26). AAPM Report No. 116 describes the pipeline behind this: raw detector data are corrected for defective pixels, offsets and gains to give “for-processing” values, and grey-scale transformations, broad-area equalization, edge restoration and noise reduction then give the “for-presentation” values that are usually stored in the picture archive (AAPM, 2009).
• Intensity relationship. Pixel Intensity Relationship says whether stored values are linear (LIN) or logarithmic (LOG) in the X-ray intensity at the detector, and Pixel Intensity Relationship Sign whether higher values mean more or less intensity; with LIN and sign −1, higher values correspond to denser structures such as bone (NEMA, 2026, C.8.11.3.1.2). Together with the grey-scale transformations, these decide whether air is displayed black or white.
• Spacing and magnification. Imager Pixel Spacing is the pixel pitch at the front of the detector and is never corrected for magnification; Pixel Spacing is used when the image has been calibrated for magnification or against an object of known size (NEMA, 2026, C.8.11.4). The DX Positioning Module records the distances from source to detector (SID) and source to patient support (SOD), noting that the latter is measured to the side of the table, support or Bucky closest to the patient, not to a point inside the patient, and the view position (NEMA, 2026, C.8.11.5).

The Digital Mammography X-Ray Image IOD specializes the DX IOD, with the same two storage classes, and adds attributes such as image laterality and whether a breast implant is present (NEMA, 2026, A.27, C.8.11.7). Computed radiography images can also be stored with the separate Computed Radiography Image IOD (NEMA, 2026, A.2), while the DX IOD lists stimulable phosphor plates among its acquisition methods (NEMA, 2026, A.26).

Characteristics

• Contrast from atomic number and density. Bone attenuates far more than soft tissue, which is why radiography’s main application is the skeleton (Berger et al., 2018), while soft tissues differ little; low tube voltages raise soft-tissue contrast, which is why mammography uses low-energy X-rays (Berger et al., 2018).
• Poisson-limited noise. Berger et al. (2018) model every stage from photon generation through attenuation to detection as a Poisson or binomial process, and the result, in the monochromatic case, is a Poisson-distributed count at the detector whose expected value is the Beer–Lambert law. The signal-to-noise ratio then grows only as the square root of the number of photons: four times the photons, and four times the dose, for twice the SNR. Their estimate neglects electronic readout noise. The model is the shot-noise part of the camera noise model in image noise, and the noise in a radiograph is accordingly signal dependent: relative to the signal, and so in the logarithmic image, it is lower behind thin regions and higher behind dense ones.
• Wide dynamic range. The X-ray intensity at the detector varies over orders of magnitude between unattenuated air and the thickest regions; storage phosphors respond over more than four of them (AAPM, 2006).
• Brightness independent of dose. In digital radiography the displayed brightness is set by processing, not by the exposure: overexposed images need not look dark, and underexposure shows as more noise rather than a light image (AAPM, 2009).

Calibration

A digital detector is calibrated so that its pixels respond alike. The corrections AAPM Report No. 116 lists for for-processing values are replacement of defective pixels, flat-field correction, per-pixel gain and offset correction, and correction of geometric distortion (AAPM, 2009). Image intensifiers need regular distortion calibration with known calibration objects, since even the Earth’s magnetic field distorts their electron optics (Berger et al., 2018).

Exposure is calibrated as well. AAPM Report No. 116 recommends that every system report an Indicated Equivalent Air Kerma, the air kerma of a standardized beam that would give the same detector response, computed from the median of for-processing values in the region of interest, and a Deviation Index,

$$\text{DI} = 10 \log_{10} \frac{K_\text{IND}}{K_\text{TGT}},$$

the deviation from a target value for the body part and view, which changes by one unit for about +25% or −20% in exposure (AAPM, 2009). The report states that the indicator reflects the exposure and noise at the detector, not the exposure to the patient.

Processing the Data

What the detector pipeline does

Between the detector and the archive, a digital radiograph is processed heavily. For storage phosphors, AAPM Report No. 93 describes the steps: the useful range of values is found from the image histogram, which depends on the anatomy and examination, and mapped to a normalized output range; a look-up table then mimics the contrast of film; and many systems apply multiscale, multifrequency processing that compresses the dynamic range and enhances contrast and edges (AAPM, 2006). Each manufacturer’s processing algorithms are proprietary, and the report observes that unacceptable image quality is often due to processing even when the raw data were acquired properly. Flat-panel systems apply similar processing to produce for-presentation images (AAPM, 2009).

Three consequences for vision models follow. Images from different vendors, or from one vendor with different processing settings, differ in contrast, edge enhancement and noise texture even for identical exposures, a source of shift between sites. The pixel values of a for-presentation image are not proportional to attenuation, so a physical model such as the one in the example applies only to for-processing data. And because the displayed brightness is normalized while noise follows the exposure, noise level and texture vary from image to image in ways tied to the patient’s size and the acquisition, information that a model can pick up whether or not it is relevant; Computer Vision in Radiology documents how acquisition differences become shortcuts.

Grids and resampling

Grid lines are a periodic pattern on the image, and digital detectors sample them. AAPM Report No. 93 gives the arithmetic: with a sampling aperture (the report’s term) of 0.2 mm the Nyquist frequency is 2.5 mm⁻¹, and a stationary grid at 3.6 mm⁻¹ is reproduced as a false pattern at 1.4 mm⁻¹ (AAPM, 2006). The report adds that displaying a reduced-size image by dropping every other pixel and row, instead of averaging, produces moiré patterns with fixed grids, which disappear only at full size. Resizing radiographs to a network’s input size is the same operation: sampling and aliasing applies, and downsampling with an anti-aliasing filter or area averaging, rather than decimation, avoids creating patterns that are not in the anatomy.

Geometry and measurement

Distances measured in pixels times Imager Pixel Spacing are distances at the detector, magnified by $M$; a structure 10 mm across at magnification 1.1 measures 11 mm. Because $M$ varies with depth, even a calibrated Pixel Spacing is exact only for the depth it was calibrated at. Registering radiographs with each other, or with CT, works with projective geometry, and medical image registration covers the methods. Projecting a CT volume through the Beer–Lambert model gives a synthetic radiograph of the same patient, a use of the physics in the example below.

Practical Example

The script simulates a radiograph of a synthetic phantom by the Beer–Lambert law with a polychromatic spectrum, then adds scatter and photon noise, using NumPy 2.3.5 and SciPy 1.16.3. Attenuation coefficients are the NIST mass attenuation coefficients for liquid water, cortical bone and soft tissue (both ICRU-44) and aluminium, with their tabulated densities, interpolated log–log between the tabulated energies (Hubbell & Seltzer, 2004). The spectrum follows Sprawls’s description of unfiltered bremsstrahlung: the photon number falls linearly to zero at 80 keV, with no characteristic lines, and 2.5 mm of aluminium filters it. The detector counts photons, each with weight one, a simplification of the energy weighting $D(E)$ that real detectors apply. These choices, the phantom, the scatter model and the photon numbers were set before the first run.

The phantom is an ellipsoid of water 20 cm thick at its centre, imaged on 256 × 256 pixels of 1 mm, with two bone rods 2 cm in diameter and a sphere of soft tissue 1 cm in diameter. Scatter is modelled crudely, as 10% of the photons removed from the primary beam reaching the detector after a Gaussian spread with a standard deviation of 40 mm. The amount is our assumption, not a measurement, and so is the resulting scatter-to-primary ratio. The script prints how the transmitted spectrum hardens with water thickness, the log contrast of a bone rod under two depths of water, and the contrast, noise and contrast-to-noise ratio (CNR) of the sphere with and without scatter, at $10^5$ and $2.5 \times 10^4$ incident photons per pixel; the contrast is the expected (noise-free) log contrast, and the noise is measured on five noise seeds. In the first run, the contrast was measured on each noisy image against a ring around the sphere, which picked up the curvature of the ellipsoid’s thickness, and the scatter-to-primary ratio was averaged over that ring; the version below compares with the same phantom without the sphere, takes the ratio behind the sphere and changes the output format. Nothing else was changed.

import numpy as np
from scipy.ndimage import gaussian_filter

# Mass attenuation coefficients mu/rho (cm^2/g) from the NIST tables
# (Hubbell & Seltzer, 2004), at 10, 15, 20, 30, 40, 50, 60, 80 keV.
E_tab = np.array([10, 15, 20, 30, 40, 50, 60, 80], float)
mu_rho = {
    "water":  [5.329, 1.673, 0.8096, 0.3756, 0.2683, 0.2269, 0.2059, 0.1837],
    "bone":   [28.51, 9.032, 4.001, 1.331, 0.6655, 0.4242, 0.3148, 0.2229],
    "tissue": [5.379, 1.699, 0.8230, 0.3790, 0.2688, 0.2264, 0.2048, 0.1823],
    "Al":     [26.23, 7.955, 3.441, 1.128, 0.5685, 0.3681, 0.2778, 0.2018],
}
rho = {"water": 1.0, "bone": 1.92, "tissue": 1.06, "Al": 2.699}  # g/cm^3

E = np.arange(10.0, 80.0, 1.0)              # keV bins; 80 kV tube
def mu(m):                                   # linear coefficient, 1/cm, log-log interpolation
    return rho[m] * np.exp(np.interp(np.log(E), np.log(E_tab), np.log(mu_rho[m])))

# Spectrum: photon number falling linearly to zero at 80 keV (a simple
# bremsstrahlung model, no characteristic lines), filtered by 2.5 mm Al.
spec = (80.0 - E) * np.exp(-mu("Al") * 0.25)
spec /= spec.sum()
E_mono = (spec * E).sum()                    # mean energy of the incident beam
mono = lambda m: np.interp(E_mono, E, mu(m))

def transmit(paths):                         # paths: {material: cm}; fraction of photons through
    att = sum(mu(m) * L for m, L in paths.items())
    return (spec * np.exp(-att)).sum(), spec * np.exp(-att)

print("incident mean energy %.1f keV" % E_mono)
print("water (cm)  mean E out (keV)  mu_eff (1/cm)  mu at %.1f keV" % E_mono)
for t in [1, 5, 10, 20]:
    f, s = transmit({"water": t})
    print("%6d %14.1f %14.3f %13.3f" % (t, (s * E).sum() / s.sum(), -np.log(f) / t, mono("water")))

print("bone rod (2 cm) log contrast:  under 5 cm water  under 20 cm water")
c = [np.log(transmit({"water": w})[0] / transmit({"water": w - 2, "bone": 2})[0]) for w in (5, 20)]
cm = (mono("bone") - mono("water")) * 2
print("  polychromatic %20.3f %18.3f\n  monoenergetic %20.3f %18.3f" % (c[0], c[1], cm, cm))

# Phantom on a 256 x 256 grid of 1 mm pixels: an ellipsoidal water body 20 cm
# thick at the centre, two 2 cm bone rods, and a 1 cm sphere of soft tissue.
y, x = np.mgrid[-127.5:128, -127.5:128]
body = 20.0 * np.sqrt(np.clip(1 - (x / 110) ** 2 - (y / 120) ** 2, 0, None))
rods = sum(0.2 * np.sqrt(np.clip(10**2 - (x - xc) ** 2, 0, None)) for xc in (-55, 55))
rods = rods * (np.abs(y) < 100)
r2 = x**2 + (y + 40) ** 2
lesion = 0.2 * np.sqrt(np.clip(5**2 - r2, 0, None))           # chord length in cm

def primary(lesion):                         # primary photons per incident photon
    att = (mu("water")[:, None, None] * (body - rods - lesion)
           + mu("bone")[:, None, None] * rods + mu("tissue")[:, None, None] * lesion)
    return np.tensordot(spec, np.exp(-att), axes=1)

def scatter(P):                              # illustrative model: 10% of the removed
    return 0.1 * gaussian_filter(1 - P, sigma=40, mode="constant")  # photons, spread widely

P1, P0 = primary(lesion), primary(0 * lesion)   # with and without the lesion
S1, S0 = scatter(P1), scatter(P0)
les, ring = r2 <= 4**2, (r2 >= 8**2) & (r2 <= 14**2)
spr = S1[les].mean() / P1[les].mean()
print("scatter-to-primary ratio behind the lesion: %.2f, 1/(1 + SPR) = %.3f" % (spr, 1 / (1 + spr)))

print("N0       scatter  contrast (%)  noise SD (%)  CNR per pixel (seeds 0-4)")
for N0 in (1e5, 2.5e4):
    for label, S, Sref in (("no", 0, 0), ("yes", S1, S0)):
        con = (np.log(P0 + Sref) - np.log(P1 + S))[les].mean()   # expected log contrast
        sds = []
        for seed in range(5):
            counts = np.random.default_rng(seed).poisson(N0 * (P1 + S))
            g = -np.log(np.maximum(counts, 1) / N0)                 # log image, bone bright
            sds.append((g + np.log(P1 + S))[ring].std())             # noise around the lesion
        sds = np.array(sds)
        print("%-8.0f %-8s %8.2f %10.2f-%.2f %10.2f-%.2f"
              % (N0, label, 100 * con, 100 * sds.min(), 100 * sds.max(),
                 con / sds.max(), con / sds.min()))

incident mean energy 44.8 keV
water (cm)  mean E out (keV)  mu_eff (1/cm)  mu at 44.8 keV
     1           46.1          0.289         0.246
     5           49.7          0.269         0.246
    10           52.6          0.256         0.246
    20           56.3          0.241         0.246
bone rod (2 cm) log contrast:  under 5 cm water  under 20 cm water
  polychromatic                1.255              0.974
  monoenergetic                1.541              1.541
scatter-to-primary ratio behind the lesion: 9.03, 1/(1 + SPR) = 0.100
N0       scatter  contrast (%)  noise SD (%)  CNR per pixel (seeds 0-4)
100000   no           1.03       2.99-3.12       0.33-0.34
100000   yes          0.10       0.96-1.01       0.10-0.11
25000    no           1.03       6.00-6.29       0.16-0.17
25000    yes          0.10       1.88-2.12       0.05-0.05

The first table is beam hardening in a projection. The transmitted spectrum’s mean energy climbs by more than 11 keV through 20 cm of water, and the effective attenuation coefficient, the log attenuation divided by the thickness, falls by about a sixth. A thin layer removes the soft end of the spectrum and looks more attenuating than water at the beam’s mean energy; a thick one transmits only the hard end and looks less. The log image is therefore not proportional to thickness, and a model of the radiograph as a sum of fixed coefficients times path lengths is only approximate.

The bone rod shows the consequence for contrast. With a monoenergetic beam its log contrast is the same whatever lies in front of it, because the water’s contribution cancels in the logarithm. With the polychromatic beam, the same rod has about a fifth less contrast under 20 cm of water than under 5 cm, since the beam reaching it is harder and bone’s advantage over water shrinks with energy. In a radiograph the intensity of a structure depends on its surroundings, one reason radiograph intensities cannot be read as tissue properties the way CT numbers can.

In the last table the sphere’s contrast with scatter is one tenth of its contrast without, the factor $1/(1 + \text{SPR})$ predicts for the scatter-to-primary ratio of about 9 that our model produces behind the sphere. The noise, measured in log units, falls with scatter, because the scattered photons add counts, yet the CNR falls by a factor close to $\sqrt{1 + \text{SPR}} \approx 3.2$, as the formula in the section on scatter predicts: scatter adds counts that carry noise but no information about the detail. Quartering the photons halves the CNR in both cases, the square-root law of Berger et al. (2018). The grid is not simulated, since grid transmission figures would add assumptions of our own; in this model it would act by lowering the scatter term.

Strengths

• A whole body region in one exposure, at a lower dose than CT; Computed Tomography quotes the FDA’s typical effective doses for a chest radiograph and a chest CT.
• High spatial resolution: flat panels have pixels of about 100–150 µm (Berger et al., 2018), and Sprawls (Blur, Resolution, and Visibility of Detail) gives blur values down to about 0.15 mm for mammography.
• Strong contrast between bone and soft tissue, which suits the skeleton (Berger et al., 2018).
• A real-time variant, fluoroscopy, for guiding procedures.

Limitations

• Superimposition: each pixel integrates attenuation through the whole body, so depth is lost and structures hide one another.
• Poor soft-tissue contrast, which tomography and other modalities address.
• Scatter and beam hardening make intensities depend on the surrounding anatomy and the acquisition; intensities are not calibrated, unlike CT numbers.
• Magnification and unsharpness vary with depth, so sizes and sharpness depend on positioning.
• Ionizing radiation. X-rays can ionize tissue and damage DNA, with a risk that increases with the number of exposures over a lifetime; NIBIB describes the cancer risk as generally considered very small and children as more sensitive (NIBIB, Medical X-rays; Berger et al., 2018).
• Exposure creep. Because digital systems produce acceptable images over a wide range of exposures, average exposures tended to rise over time after departments moved from screen-film to digital imaging, the problem that motivated AAPM Report No. 116 (AAPM, 2009).

Variants

Fluoroscopy acquires a sequence of radiographs at a frame rate, with image intensifiers or flat panels; rates of 30 frames per second are possible but are often reduced for dose, and flat-panel frame rates typically range from 7.5 to 30 per second, with binning of 2 × 2 or 4 × 4 pixels raising the rate (Berger et al., 2018). It guides minimally invasive procedures, where catheters and other tools must be steered without a direct view, often with a C-arm that can be positioned freely around the patient. Image intensifiers bring their own artifacts, vignetting and distortion. Digital subtraction angiography subtracts a mask image acquired before an iodine contrast agent is injected from images acquired after it, leaving only the vessels, provided the patient has not moved (Berger et al., 2018). Motion between mask and fill images is a registration problem, the kind image registration addresses.

Mammography images the breast compressed between a detector support and a paddle (NIBIB, Medical X-rays), with low-energy spectra from molybdenum or rhodium anodes, filters that use the molybdenum K-edge to remove the higher-energy bremsstrahlung, and small focal spots (Sprawls, X-ray Production; Berger et al., 2018). Breast density affects how visible abnormalities are (NIBIB, Medical X-rays). Its images use their own DICOM IOD, and its screening application, with multiple views and prior examinations, is the subject of mammography and breast cancer screening.

Typical Applications

Berger et al. (2018) name the skeleton as radiography’s main application area, fractures and changes in bone density among them, since bone contrasts well with surrounding tissue. Chest radiographs are the subject of most large public datasets in medical vision, as Computer Vision in Radiology describes; chest X-ray analysis covers their findings, labels and datasets. Mammography is read in screening programmes. Fluoroscopy and angiography guide interventions. Outside medicine, X-rays inspect metal parts for small cracks in non-destructive testing (Berger et al., 2018).

History

Wilhelm Conrad Röntgen discovered X-rays in Würzburg on 8 November 1895, while experimenting with a Crookes tube covered in black cardboard, when he saw a fluorescent screen glow at a distance; he submitted his first article, “Über eine neue Art von Strahlen”, on 28 December 1895, and demonstrated the rays in January 1896 by imaging the hand of the anatomist Albert von Kölliker (Berger et al., 2018). He received the first Nobel Prize in Physics in 1901 for his discovery of “the remarkable rays subsequently named after him” (NobelPrize.org, Röntgen facts).

For most of the following century the receptor was film behind intensifying screens, which reduced the exposure film alone would need (Sprawls, Screen/Film Radiographic Receptors). Fluoroscopy first viewed a single phosphor screen, so dim that radiologists had to adapt their eyes in the dark; image intensifiers, introduced in the late 1940s, made the brightness adjustable (Berger et al., 2018). Storage-phosphor plates brought digital images within the film workflow, using cassettes that look like screen-film cassettes (AAPM, 2006), and flat-panel detectors, introduced in the mid-1990s, read the image out directly (Berger et al., 2018). Digitization separated acquisition from display, which brought wide dynamic range and image processing, and also exposure creep and the push for standard exposure indicators (AAPM, 2009).

Comparison with Other Sensors

• Computed tomography measures many projections and reconstructs calibrated cross-sections, removing superimposition at a higher dose; Computed Tomography covers reconstruction and artifacts, and Hounsfield Units covers its calibrated scale.
• Magnetic resonance imaging uses no ionizing radiation and gives cross-sectional soft-tissue contrast, with slower acquisitions, as the MRI article describes.
• Ultrasound uses no ionizing radiation but images bone and air-filled tissue such as the lungs poorly (NIBIB, Ultrasound); Ultrasound Imaging describes the sensor.
• Cameras record light reflected from surfaces under a lens; a radiograph records transmission through the object along rays from a point source, so it sees inside objects, and its pixel values sum along depth instead of describing the nearest surface.