Compton Scattering Tomography Overview

• CST is an imaging modality that reconstructs 2D electron density by measuring Compton scattered photons using energy-resolved data.
• It encompasses a range of inverse problems, employing various forward operators like disc transforms, circular arcs, and Radon-type operators.
• Advanced reconstruction techniques in CST address challenges such as multiple scattering, attenuation, and artifacts via analytic, numerical, and learning-based methods.

Compton Scattering Tomography (CST) is an imaging modality that seeks to reconstruct the two-dimensional electron density of a slice through an object by using measurements of photons scattered via the Compton effect. In the broader literature, CST denotes a family of inverse problems in which energy-resolved scatter measurements are mapped to electron density, attenuation, mass density, photoelectric response, contours, or support, depending on the acquisition geometry and physical model. The associated forward operators include disc integrals, circular-arc and double-arc transforms, toric-section transforms, spindle and spindle-interior transforms, apple and lemon surface transforms, V-line transforms, and weighted or non-linear Radon-type operators (Webber, 2015, Webber et al., 2017, Webber et al., 2019, Tarpau et al., 2019, Tarpau et al., 2020, Webber et al., 26 Jul 2025).

1. Physical basis and measurement model

At the core of CST lies the Compton energy–angle relation. For an incident photon of energy $E_0$ scattering through angle $\theta$, the outgoing energy $E'$ satisfies

$E'=\frac{E_0}{1+\frac{E_0}{m_ec^2}(1-\cos\theta)},$

with $m_ec^2 \simeq 511\,\mathrm{keV}$. In practice, an energy-resolving detector measures the scattered energy and thereby determines the scattering angle. Because the probability of Compton scattering is proportional to the local electron density $n_e(x)$, the scattered count rate encodes weighted integrals of electron density over geometry-dependent manifolds (Tarpau et al., 2020).

For first-order scattering, a typical local intensity model takes the form

$$dI_1 =F(x)\;n_e(x)\;\frac{d\sigma_c}{d\Omega}(E_0,\theta)\;d\Omega\;dV,$$

with the Klein–Nishina differential cross section

$$\frac{d\sigma_c}{d\Omega}(E_0,\theta) =\frac{r_0^2}{2}\Bigl(\tfrac{E'}{E_0}\Bigr)^2 \Bigl(\tfrac{E'}{E_0}+\tfrac{E_0}{E'}-\sin^2\theta\Bigr).$$

In x-ray CST with a polychromatic source, the recorded signal is additionally modulated by attenuation along incident and scattered paths, source flux, detector solid angle, finite detector aperture, and detector energy resolution (Webber, 2015, Tarpau et al., 2020).

A recurrent simplification is the single-scatter approximation. In the early 2-D x-ray disc-transform formulation, dark-field measurements are assumed to be dominated by the Compton scattering process, so that singly-scattered photons contribute appreciably while the forward model reduces to integrals over discs whose boundaries intersect the source point (Webber, 2015). Later work emphasizes that this approximation is not generally faithful to measured spectral data: second- and higher-order scattering are a substantial part of the spectral measurement, and it is in general impossible to know how many times a photon was scattered before being measured (Kuger et al., 2020). This marks one of the central modeling tensions in CST.

2. Acquisition geometries and associated integral transforms

CST is not a single scanner architecture. The literature contains several geometries, each with a distinct data manifold and inverse problem.

Geometry	Integration manifold	Representative result
2-D x-ray dark-field slice	discs through source point	disc transform reduction to Radon (Webber, 2015)
Circular CST / CCST	circular arcs	weighted Radon formulation on arcs (Tarpau et al., 2020)
Uncollimated rotating detector	double circular arcs	analytic inversion formula (Tarpau et al., 2019)
Ring geometry for baggage screening	toric sections	injective, but $2\!-\!1$ canonical relation (Webber et al., 2019)
3-D monochromatic CST	spindle tori	injectivity on a hollow ball (Webber et al., 2017)
3-D fixed-source modality	toric “apple” surfaces	generalized Abel-type inversion (Cebeiro et al., 2020)
3-D cylindrical scanner	lemon surfaces	injective lemon transform and Bolker analysis (Webber, 2023)

In the 2-D x-ray model with fixed energy-sensitive detectors and a polychromatic source, the locus of possible scatterers for a given source point and dark-field detector is approximately a planar disc $D$, even though the exact manifold is a toric section (Webber, 2015). In circular CST with a monochromatic source at the origin and a detector ring, each energy bin corresponds to the union of two circular arcs passing through source and detector, and the forward operator can be written as a weighted Radon transform on circular arcs (Tarpau et al., 2020).

Several works study non-collimated systems. In the uncollimated single-detector geometry, each energy corresponds to two circular arcs, and the data define a double-arc Radon transform with a closed-form inversion obtained by circular harmonic expansion and Cormack-type consistency (Tarpau et al., 2019). In a translational geometry intended for conveyor-style scanning, backscattered photons generate toric-section curves in 2-D and apple surfaces in 3-D, leading again to Volterra-type inversion problems (Webber et al., 2019, Tarpau et al., 2021).

Three-dimensional CST introduces surfaces of revolution. Monochromatic source–detector motion on the sphere yields spindle tori; fixed-source, moving-detector systems generate toric “apple” surfaces; cylindrical scanners generate “lemons”; and full seven-dimensional analyses examine apple and lemon surface families as Fourier Integral Operators (FIOs) (Webber et al., 2017, Cebeiro et al., 2020, Webber et al., 2022, Webber, 2023). A plausible implication is that CST should be understood less as a single transform than as a class of geometry-dependent generalized Radon transforms tied together by Compton kinematics.

3. Inversion theory in two dimensions

The foundational 2-D x-ray paper formulates a disc integral operator

$\theta$0

or equivalently $\theta$1, where the disc boundary passes through the current source point (Webber, 2015). The key analytical step is the introduction of the involution

$\theta$2

for which

$\theta$3

with $\theta$4 the standard 2-D Radon transform. This reduces the disc problem to ordinary Radon inversion. If $\theta$5 is smooth and supported in the annulus $\theta$6, then the full data determine $\theta$7 uniquely, hence $\theta$8 uniquely (Webber, 2015).

The same work gives an explicit inversion formula by differentiating the disc data, applying filtered back-projection to $\theta$9, and inverting the involution. It also proves Sobolev estimates showing that $E'$0 is continuous $E'$1 for $E'$2, and that the inverse is mildly ill-posed according to Natterer’s criterion. For finite $E'$3-sampling with Hausdorff-gap $E'$4, the estimate $E'$5 quantifies the sampling error (Webber, 2015).

Other 2-D geometries exhibit different inversion structures. In the uncollimated rotating-detector setting, the double circular arc transform admits a closed-form inversion

$E'$6

together with a Hilbert-transform implementation that is fast and efficient numerically (Tarpau et al., 2019). By contrast, the static circular geometry studied in non-collimated circular CST is described in discrete form by $E'$7, and no closed-form inversion for the double-arc transform $E'$8 is known today because $E'$9 lacks shift and rotation invariance (Ayad et al., 2024).

Microlocal analysis reveals that invertibility does not preclude artifacts. For the toric-section transform in a ring geometry, each branch $E'=\frac{E_0}{1+\frac{E_0}{m_ec^2}(1-\cos\theta)},$0 or $E'=\frac{E_0}{1+\frac{E_0}{m_ec^2}(1-\cos\theta)},$1 satisfies the Bolker Assumption, but the full canonical relation is $E'=\frac{E_0}{1+\frac{E_0}{m_ec^2}(1-\cos\theta)},$2. The result is a predictable family of nonlocal image artifacts in filtered backprojection; nevertheless, the forward operator is injective for $E'=\frac{E_0}{1+\frac{E_0}{m_ec^2}(1-\cos\theta)},$3 functions supported inside the open unit ball (Webber et al., 2019). This distinction between injectivity and artifact-free inversion recurs throughout CST.

4. Three-dimensional CST and microlocal structure

In three dimensions, the forward operators typically reduce modewise to Volterra or generalized Abel equations after harmonic decomposition. For spindle-torus data generated by a monochromatic source, injectivity is proved on smooth functions compactly supported on a hollow ball by expanding $E'=\frac{E_0}{1+\frac{E_0}{m_ec^2}(1-\cos\theta)},$4 in spherical harmonics and reducing the problem to explicitly invertible Volterra operators (Webber et al., 2017). Polychromatic variants lead to spindle-interior, apple, and apple-interior transforms, again injective under the stated support assumptions (Webber et al., 2017).

A fixed-source, moving-detector modality on a spherical surface leads to a toric Radon transform on “apple” tori. In this case the data coefficients $E'=\frac{E_0}{1+\frac{E_0}{m_ec^2}(1-\cos\theta)},$5 satisfy a generalized Abel-type equation in the radial variable, and uniqueness follows from weakly singular Volterra theory. Numerical inversion is performed modewise with Tikhonov regularization, because $E'=\frac{E_0}{1+\frac{E_0}{m_ec^2}(1-\cos\theta)},$6 can be ill-conditioned when $E'=\frac{E_0}{1+\frac{E_0}{m_ec^2}(1-\cos\theta)},$7 is small or vanishes. On a $E'=\frac{E_0}{1+\frac{E_0}{m_ec^2}(1-\cos\theta)},$8 phantom, the reported errors were $E'=\frac{E_0}{1+\frac{E_0}{m_ec^2}(1-\cos\theta)},$9 and $m_ec^2 \simeq 511\,\mathrm{keV}$0 in the noiseless run, and $m_ec^2 \simeq 511\,\mathrm{keV}$1, $m_ec^2 \simeq 511\,\mathrm{keV}$2 at $m_ec^2 \simeq 511\,\mathrm{keV}$3 noise (Cebeiro et al., 2020).

The cylindrical 3-D scanner introduces a lemon Radon transform $m_ec^2 \simeq 511\,\mathrm{keV}$4 and a fixed-distance restricted transform $m_ec^2 \simeq 511\,\mathrm{keV}$5. Fourier decomposition in the axial and angular variables yields Volterra equations of the first kind with smooth kernels $m_ec^2 \simeq 511\,\mathrm{keV}$6, from which injectivity of both $m_ec^2 \simeq 511\,\mathrm{keV}$7 and $m_ec^2 \simeq 511\,\mathrm{keV}$8 on $m_ec^2 \simeq 511\,\mathrm{keV}$9 is derived (Webber, 2023). The same paper proves that $n_e(x)$0 satisfies the Bolker condition and that $n_e(x)$1 is a pseudodifferential operator, so visible singularities are mapped stably to data and back. However, because the lemons sweep only a restricted set of angles, edges whose normals lie close to the cylinder axis may be invisible; reconstructions without strong regularization therefore exhibit blurring along the $n_e(x)$2-direction (Webber, 2023).

The seven-dimensional microlocal study of apple and lemon transforms sharpens this picture. In the full seven-dimensional case, both transforms are elliptic FIOs. Under the support restriction $n_e(x)$3, the lemon transform satisfies the Bolker condition, whereas reduced-data apple transforms violate it and generate artifacts on apple-cylinder intersections (Webber et al., 2022). This shows that stability in CST depends not only on the physics of scattering but also on whether the acquisition family preserves the injective-immersion structure of the left projection.

5. Multiple scattering, attenuation, and multimodal recovery

A common misconception is that scatter is merely a nuisance to be removed. In conventional modalities this is true, but CST explicitly uses the Compton effect to image electron density, and in several works the same acquisition platform also records transmitted photons for attenuation imaging (Tarpau et al., 2020). A second misconception is that first-order scattering alone adequately models spectral data. The later multiple-scattering analysis states the opposite: second- and higher-order scattering are a substantial part of the spectral measurement (Kuger et al., 2020).

The first-order term can be written as a weighted circular Radon transform

$n_e(x)$4

where the weight contains the Klein–Nishina factor, geometric decay, and attenuation (Kuger et al., 2020). Second-order scattering has a double-integral representation over $n_e(x)$5, with two Klein–Nishina factors and three attenuation terms, and higher orders are conjectured to form progressively smoother generalized Radon-type operators (Kuger et al., 2020). In a related 3-D contour-reconstruction study, first- and second-order data are modeled explicitly, and microlocal analysis shows that the second-order component is structurally smoother than the first-order component, so the contours of the electron density are essentially encoded within the first-order part (Rigaud, 2019).

This smoothing hierarchy motivates differentiated reconstruction strategies. In a joint CST–CT fan-beam scanner, one first reconstructs a sparse-view CT prior $n_e(x)$6, then applies a regularized differentiation $n_e(x)$7 in energy to accentuate the half-order singularities of $n_e(x)$8 while nearly annihilating the smoother $n_e(x)$9. The resulting differentiated CST plus TV scheme resolves a $$dI_1 =F(x)\;n_e(x)\;\frac{d\sigma_c}{d\Omega}(E_0,\theta)\;d\Omega\;dV,$$0 crack in an aluminium ring and remains superior to direct first-order inversion when $$dI_1 =F(x)\;n_e(x)\;\frac{d\sigma_c}{d\Omega}(E_0,\theta)\;d\Omega\;dV,$$1 is present (Kuger et al., 2020).

Attenuation can also be exploited constructively. In the 2-D disc-transform formulation, combining light-field attenuation reconstruction with dark-field electron-density recovery yields pointwise identification of the effective atomic number $$dI_1 =F(x)\;n_e(x)\;\frac{d\sigma_c}{d\Omega}(E_0,\theta)\;d\Omega\;dV,$$2 under the assumptions that $$dI_1 =F(x)\;n_e(x)\;\frac{d\sigma_c}{d\Omega}(E_0,\theta)\;d\Omega\;dV,$$3 and $$dI_1 =F(x)\;n_e(x)\;\frac{d\sigma_c}{d\Omega}(E_0,\theta)\;d\Omega\;dV,$$4 is known, continuous, and strictly increasing in $$dI_1 =F(x)\;n_e(x)\;\frac{d\sigma_c}{d\Omega}(E_0,\theta)\;d\Omega\;dV,$$5 for $$dI_1 =F(x)\;n_e(x)\;\frac{d\sigma_c}{d\Omega}(E_0,\theta)\;d\Omega\;dV,$$6 up to approximately $$dI_1 =F(x)\;n_e(x)\;\frac{d\sigma_c}{d\Omega}(E_0,\theta)\;d\Omega\;dV,$$7 (Webber, 2015). In circular CST design, a bi-imaging configuration is proposed in which transmitted photons recover $$dI_1 =F(x)\;n_e(x)\;\frac{d\sigma_c}{d\Omega}(E_0,\theta)\;d\Omega\;dV,$$8 and scattered photons recover $$dI_1 =F(x)\;n_e(x)\;\frac{d\sigma_c}{d\Omega}(E_0,\theta)\;d\Omega\;dV,$$9, with alternating or coupled inversion such as Landweber or Kaczmarz iterations (Tarpau et al., 2020). In a limited-view fusion framework, a joint variational model combines attenuation and Compton scatter data to recover mass density $$\frac{d\sigma_c}{d\Omega}(E_0,\theta) =\frac{r_0^2}{2}\Bigl(\tfrac{E'}{E_0}\Bigr)^2 \Bigl(\tfrac{E'}{E_0}+\tfrac{E_0}{E'}-\sin^2\theta\Bigr).$$0 and photoelectric coefficient $$\frac{d\sigma_c}{d\Omega}(E_0,\theta) =\frac{r_0^2}{2}\Bigl(\tfrac{E'}{E_0}\Bigr)^2 \Bigl(\tfrac{E'}{E_0}+\tfrac{E_0}{E'}-\sin^2\theta\Bigr).$$1; on the reported $$\frac{d\sigma_c}{d\Omega}(E_0,\theta) =\frac{r_0^2}{2}\Bigl(\tfrac{E'}{E_0}\Bigr)^2 \Bigl(\tfrac{E'}{E_0}+\tfrac{E_0}{E'}-\sin^2\theta\Bigr).$$2 phantoms, fused reconstructions reduced mass-density RMSE from $$\frac{d\sigma_c}{d\Omega}(E_0,\theta) =\frac{r_0^2}{2}\Bigl(\tfrac{E'}{E_0}\Bigr)^2 \Bigl(\tfrac{E'}{E_0}+\tfrac{E_0}{E'}-\sin^2\theta\Bigr).$$3 or $$\frac{d\sigma_c}{d\Omega}(E_0,\theta) =\frac{r_0^2}{2}\Bigl(\tfrac{E'}{E_0}\Bigr)^2 \Bigl(\tfrac{E'}{E_0}+\tfrac{E_0}{E'}-\sin^2\theta\Bigr).$$4 to $$\frac{d\sigma_c}{d\Omega}(E_0,\theta) =\frac{r_0^2}{2}\Bigl(\tfrac{E'}{E_0}\Bigr)^2 \Bigl(\tfrac{E'}{E_0}+\tfrac{E_0}{E'}-\sin^2\theta\Bigr).$$5 for Phantom I and from $$\frac{d\sigma_c}{d\Omega}(E_0,\theta) =\frac{r_0^2}{2}\Bigl(\tfrac{E'}{E_0}\Bigr)^2 \Bigl(\tfrac{E'}{E_0}+\tfrac{E_0}{E'}-\sin^2\theta\Bigr).$$6 or $$\frac{d\sigma_c}{d\Omega}(E_0,\theta) =\frac{r_0^2}{2}\Bigl(\tfrac{E'}{E_0}\Bigr)^2 \Bigl(\tfrac{E'}{E_0}+\tfrac{E_0}{E'}-\sin^2\theta\Bigr).$$7 to $$\frac{d\sigma_c}{d\Omega}(E_0,\theta) =\frac{r_0^2}{2}\Bigl(\tfrac{E'}{E_0}\Bigr)^2 \Bigl(\tfrac{E'}{E_0}+\tfrac{E_0}{E'}-\sin^2\theta\Bigr).$$8 for Phantom II (Rezaee et al., 2017).

Recent non-linear analysis treats attenuation as part of the forward operator rather than a frozen weight. The non-linear Compton transform

$$\frac{d\sigma_c}{d\Omega}(E_0,\theta) =\frac{r_0^2}{2}\Bigl(\tfrac{E'}{E_0}\Bigr)^2 \Bigl(\tfrac{E'}{E_0}+\tfrac{E_0}{E'}-\sin^2\theta\Bigr).$$9

uses a V-line transform to model incoming and outgoing attenuation. Its strongest singularities are shown to coincide with those of the classical weighted Radon transform $2\!-\!1$0, and this yields support and boundary recovery theorems as well as a uniqueness result for constant density under the stated assumptions on $2\!-\!1$1 and the smoothing kernel (Webber et al., 26 Jul 2025). This suggests that non-linearity in CST does not erase the dominant microlocal content of the underlying Radon geometry, even though it complicates exact inversion.

6. Computation, learning-based reconstruction, applications, and open problems

Classical CST reconstruction pipelines are typically hybrid analytic–numerical schemes. In the disc-transform setting the sequence is: preprocessing by dividing out a geometric–attenuation–cross-section factor, smoothing in the disc parameter, finite-difference differentiation, filtered back-projection using a windowed ramp filter such as Hamming, involution inversion, and multi-view averaging (Webber, 2015). In simulated 2-D water-bottle data, the reported average reconstruction error was approximately $2\!-\!1$2 at $2\!-\!1$3 noise, approximately $2\!-\!1$4 at $2\!-\!1$5 noise, and up to approximately $2\!-\!1$6 at $2\!-\!1$7 noise, while average recovered pixel values remained $2\!-\!1$8–$2\!-\!1$9 for a true value of $D$0 (Webber, 2015).

System design papers emphasize that reconstruction quality is inseparable from detector physics. In circular CST, coarse energy resolution $D$1 removes data for $D$2 and produces central artifacts; finite detector size turns a single arc into a circular sector and increases blur; and imperfect collimation or higher-order scatter introduces background (Tarpau et al., 2020). In a detector concept optimized for hard-x-ray Compton imaging, a $D$3-thick xenon electroluminescent TPC was simulated to process photon rates up to at least $D$4 on-sample, with approximately $D$5 detection efficiency at $D$6; using the Rose criterion and the dose partitioning theorem, the same study projected $D$7 resolution in about $D$8 for $D$9-size unstained cells (Hernández et al., 2020).

Data-driven methods have entered CST primarily where analytic inversion is unavailable or unstable. In the static non-collimated circular geometry, UnWave-Net unrolls

$\theta$00

with wavelet-based non-local regularization and a Swin Transformer acting only on the low-frequency LL subband (Ayad et al., 2024). On $\theta$01 clinical CT slices forward projected into CST data, the reported no-noise, $\theta$02 performance was PSNR $\theta$03 and SSIM $\theta$04, exceeding pseudoinverse, TV, U-Net, DuDoTrans, LPD, LEARN, and RegFormer; the same method used $\theta$05M parameters, approximately $\theta$06 memory, and $\theta$07 inference time on a single V100 (Ayad et al., 2024).

A second computational direction addresses forward-model uncertainty directly. In Monte-Carlo-based CST with exact and approximate forward operators, RESESOP treats operator and data errors through stripe projections, while Deep Image Prior incorporates the same uncertainty into an unsupervised loss. On a modified Shepp–Logan setup with $\theta$08 source–detector pairs and $\theta$09 energies, RESESOP+TV achieved SSIM $\theta$10, PSNR $\theta$11, and NMSE $\theta$12 in the exact $\theta$13 scenario, while DIP with the RESESOP-inspired loss attained SSIM $\theta$14 and PSNR $\theta$15 (Gödeke et al., 2022). These results indicate that model uncertainty is not peripheral in CST; it is often the central computational constraint.

The application domains named in the literature include industrial objects scanned from one side, cultural heritage, security and baggage scanning, non-destructive testing, and biomedical niche uses such as lung imaging under therapy (Tarpau et al., 2020). Yet the principal open problems remain fundamental: analytic inversion of the fully weighted, attenuated Radon transform on arcs is still open; full 3-D extensions lead to toroidal or other generalized surface transforms whose inversion is more complex or largely unstudied; multiple scattering, finite energy resolution, detector blur, and attenuation-model mismatch continue to limit quantitative accuracy; and limited-angle effects create invisible singularities or predictable artifact families when the Bolker condition fails (Tarpau et al., 2020, Webber et al., 2022, Webber, 2023). The cumulative record therefore presents CST not as a single settled method, but as a mathematically rich and still evolving domain in inverse problems, microlocal analysis, detector design, and computational imaging.