Spherical Region of Interest (SRoI)

• SRoI is a spherical region defined for targeted data acquisition, reconstruction, and analysis, offering rotational symmetry and analytical tractability.
• It underpins methodologies in cone-beam CT, interior tomography, spatial acoustics, and free-space turbulence by localizing the area of interest and managing incomplete data.
• SRoI brings unique challenges such as truncated measurements and boundary-induced artifacts, motivating specialized inversion algorithms and deep-learning corrections.

Searching arXiv for papers related to “Spherical Region of Interest (SRoI)” and closely related terminology across domains. A Spherical Region of Interest (SRoI) is a sphere or ball-shaped subdomain selected for targeted reconstruction, analysis, control, or simulation within a larger physical or computational domain. Across the literature, the term is used most explicitly in ROI tomography for cone-beam CT, where only rays intersecting a spherical subset $C\subset B$ are acquired and used for reconstruction (Azencott et al., 2015). Closely related formulations appear in interior tomography with circular or disk-like ROIs (Han et al., 2017), in free-space turbulence where an initially spherical compact turbulent cloud functions as a localized region of interest embedded in irrotational surroundings (Yu et al., 2020), and in spatial acoustics where the controlled listening area is a spherical region rather than a set of discrete sweet spots (Wang et al., 2023). A plausible synthesis is that SRoI denotes a geometry-driven localization strategy: the sphere is used because it is analytically tractable, rotationally symmetric, and compatible with basis expansions, inverse operators, and interface analysis. At the same time, the literature shows that the implications of an SRoI are strongly domain-specific, ranging from data truncation and null-space ambiguity in imaging to edge intermittency and vortex-ring ejection in turbulence.

1. Conceptual scope and geometric definition

In the most explicit formulation, an SRoI is a spherical ROI $C\subset B$ inside a bounded ball $B\subset \mathbb{R}^3$, with reconstruction based only on the subset of cone-beam rays intersecting $C$ (Azencott et al., 2015). In that setting, the non-truncated cone-beam transform is

$Df(a,\theta)=\int_0^\infty f(a+t\theta)\,dt,$

while the ROI-truncated transform is the restriction

$D_C f = 1_{\mathcal R_C} Df,$

where $\mathcal R_C\subset \mathcal R_B$ is the set of rays intersecting the spherical ROI (Azencott et al., 2015). The SRoI is therefore not merely a visual crop; it is the subset that determines which measurements are acquired.

In interior tomography, the corresponding geometry is two-dimensional: the ROI is disk-like, and the truncated Radon transform is restricted to detector coordinates $|s|<\mu$ (Han et al., 2017). The paper does not formalize an SRoI in 3D terms, but it identifies the circular ROI as the natural two-dimensional analogue of a spherical ROI. This suggests a general geometric principle: the “spherical” descriptor refers less to a particular application than to a rotationally symmetric compact support within which one seeks accurate inference under incomplete external data.

Outside tomography, the same geometric motif appears in physically different ways. In turbulence, the initial condition is a compact, windowed sphere of isotropic homogeneous incompressible turbulence embedded in otherwise quiescent free space, created using

$$\Phi(r)=\frac{1}{2}\left[1-\tanh\left(\frac{2(r-R)}{\sigma}\right)\right]$$

to localize the vorticity field (Yu et al., 2020). In acoustics, the listening region is modeled as a source-free spherical region in which acoustic velocity vectors are reproduced throughout the full volume, not just on a boundary or at a few control points (Wang et al., 2023). In 360° imaging, the literature does not define an SRoI formally, but local spherical faces and their neighborhoods are used as the operative spherical subregions for implicit decoding on $S^2$ (Yoon et al., 2021). In wide-field radio interferometry, the closest analogue is a spherical field of view on $C\subset B$0, restricted to a cap of angular radius $C\subset B$1, rather than a named SRoI (Tajja et al., 3 Mar 2025).

2. SRoI in cone-beam CT and interior tomography

ROI tomography addresses reconstruction of a region of interest using only projections that intersect that region, with the stated goal of reducing overall radiation exposure when only a small specific region must be examined (Azencott et al., 2015). For an unknown density

$C\subset B$2

the acquisition geometry places sources on a smooth curve $C\subset B$3 satisfying the classical strong Tuy condition

$C\subset B$4

for every point $C\subset B$5 and direction $C\subset B$6 (Azencott et al., 2015). Under full data, an explicit inverse $C\subset B$7 exists; under ROI truncation, direct application of $C\subset B$8 to $C\subset B$9 is not valid (Azencott et al., 2015).

The proposed reconstruction strategy defines

$B\subset \mathbb{R}^3$0

with smoothing operators $B\subset \mathbb{R}^3$1, initial term

$B\subset \mathbb{R}^3$2

and fixed-point iteration

$B\subset \mathbb{R}^3$3

The limit

$B\subset \mathbb{R}^3$4

defines an approximate inverse of $B\subset \mathbb{R}^3$5 when the sphere $B\subset \mathbb{R}^3$6 is sufficiently large (Azencott et al., 2015). The convergence mechanism is controlled by the estimate

$B\subset \mathbb{R}^3$7

which leads to

$B\subset \mathbb{R}^3$8

If the ROI is sufficiently large relative to $B\subset \mathbb{R}^3$9, $C$0 becomes a contraction and the iteration converges exponentially (Azencott et al., 2015).

The reconstruction guarantee is formulated through the notion of an $C$1-accurate inverse: $C$2 The paper proves the existence of a critical ROI radius $C$3 such that $C$4 ensures convergence and $C$5-accurate approximation (Azencott et al., 2015). Numerical experiments on spherical, spiral, C-arm, and twin-circle acquisition geometries, with spherical ROIs of radii $C$6 voxels and full radius about $C$7 voxels, reported relative $C$8 error inside the ROI typically in the range of about $C$9–$Df(a,\theta)=\int_0^\infty f(a+t\theta)\,dt,$0, with estimated critical radii for $Df(a,\theta)=\int_0^\infty f(a+t\theta)\,dt,$1 relative error roughly $Df(a,\theta)=\int_0^\infty f(a+t\theta)\,dt,$2–$Df(a,\theta)=\int_0^\infty f(a+t\theta)\,dt,$3 voxels for spherical setups, $Df(a,\theta)=\int_0^\infty f(a+t\theta)\,dt,$4–$Df(a,\theta)=\int_0^\infty f(a+t\theta)\,dt,$5 voxels for spiral, and $Df(a,\theta)=\int_0^\infty f(a+t\theta)\,dt,$6–$Df(a,\theta)=\int_0^\infty f(a+t\theta)\,dt,$7 voxels for circle and twin-circle cases (Azencott et al., 2015).

A distinct but related issue appears in deep-learning interior tomography. There, the truncated Radon transform has a nontrivial null space,

$Df(a,\theta)=\int_0^\infty f(a+t\theta)\,dt,$8

which manifests as cupping artifacts in FBP reconstructions from truncated data (Han et al., 2017). The proposed remedy is an FBP + CNN pipeline that treats FBP as a practical right inverse and trains a modified U-Net to map artifact-corrupted reconstructions to ground truth ROI images (Han et al., 2017). On AAPM Low-Dose CT data, using only the central 350 of 736 detectors and a $Df(a,\theta)=\int_0^\infty f(a+t\theta)\,dt,$9 center ROI crop, the reported metrics were PSNR 37.4600 dB for the proposed method, 30.2004 dB for TV, 27.0344 dB for the Lee method, and 9.4099 dB for FBP, with per-slice runtimes of 0.0532 s, 1.8272 s, and 0.3438 s, respectively (Han et al., 2017). This establishes a second SRoI-related line of work in which the sphere-like ROI induces an inverse problem whose primary pathology is null-space ambiguity rather than insufficient sampling density alone.

3. Numerical and analytical structures induced by spherical localization

A recurrent feature of SRoI formulations is that spherical localization changes both the admissible operators and the dominant error mechanisms. In cone-beam CT, truncation replaces an exactly invertible full operator $D_C f = 1_{\mathcal R_C} Df,$0 by a restricted operator $D_C f = 1_{\mathcal R_C} Df,$1, so the missing-data operator

$D_C f = 1_{\mathcal R_C} Df,$2

becomes central to both theory and algorithm design (Azencott et al., 2015). The key fact is that the geometry of a sufficiently large sphere makes the missing set small enough for contraction-based correction after regularization.

In interior tomography, the principal mathematical structure is the null space of the truncated Radon transform, not merely loss of stability in a generic sense (Han et al., 2017). The CNN is therefore interpreted as learning a projection that removes components $D_C f = 1_{\mathcal R_C} Df,$3 from an FBP reconstruction of the form $D_C f = 1_{\mathcal R_C} Df,$4, with the intended mapping

$D_C f = 1_{\mathcal R_C} Df,$5

(Han et al., 2017). This suggests that SRoI imaging problems often require explicit treatment of measurement incompleteness induced by geometry, rather than treating truncation as a conventional denoising problem.

In spatial acoustics, the spherical region supports a spherical-harmonic representation that is volumetric rather than pointwise. Pressure in the spherical listening region is expanded as

$D_C f = 1_{\mathcal R_C} Df,$6

and the local acoustic velocity vectors are obtained from translated local coefficients $D_C f = 1_{\mathcal R_C} Df,$7, with only $D_C f = 1_{\mathcal R_C} Df,$8 terms contributing to velocity at the evaluation point (Wang et al., 2023). The resulting AVV field is represented through radial-independent SHV coefficients

$D_C f = 1_{\mathcal R_C} Df,$9

so that one coefficient set characterizes the AVVs throughout the whole spherical region (Wang et al., 2023). The sphere here is not a truncation device but the natural domain on which a basis expansion acquires region-wide meaning.

In turbulence, spherical localization imposes exact free-space boundary conditions on a compact vortical region through the fast lattice Green’s function method on $\mathcal R_C\subset \mathcal R_B$0 (Yu et al., 2020). The SRoI-like sphere is thus embedded in a surrounding flow that is physically active and not discarded. The solver truncates the computational work to a finite, adaptively moving region around the turbulence while retaining the infinite-domain physics exactly for compact vorticity fields (Yu et al., 2020). A plausible implication is that SRoI formulations are particularly effective when the spherical geometry can be paired with an operator whose analytic structure respects the exterior domain rather than replacing it with artificial boundaries.

4. Free-space turbulence as an SRoI laboratory

The paper on the dynamics and decay of a spherical region of turbulence in free space studies an initially spherical compact turbulent region evolving in otherwise quiescent free space (Yu et al., 2020). The initial condition is formed from isotropic homogeneous incompressible turbulence taken from a periodic DNS, tiled into free space, and multiplied by the spherical window $\mathcal R_C\subset \mathcal R_B$1, producing a sphere of nonzero vorticity of radius $\mathcal R_C\subset \mathcal R_B$2 and transition width $\mathcal R_C\subset \mathcal R_B$3 (Yu et al., 2020). Deep inside the sphere, the turbulence is approximately locally homogeneous and isotropic; near the boundary, a turbulent/non-turbulent interface develops (Yu et al., 2020).

A major methodological component is the fast lattice Green’s function method for the incompressible Navier–Stokes equations on an effectively infinite grid, coupled with third-order Runge–Kutta time stepping and an integrating factor treatment of viscosity (Yu et al., 2020). The spatially adaptive grid uses blocks of $\mathcal R_C\subset \mathcal R_B$4 cells, DNS uses up to about $\mathcal R_C\subset \mathcal R_B$5 cells on 1500 cores, and LES employs the stretched vortex sub-grid stress model, with SGS tensor

$\mathcal R_C\subset \mathcal R_B$6

and stretched-spiral vortex spectrum

$\mathcal R_C\subset \mathcal R_B$7

(Yu et al., 2020).

The initial low-wavenumber behavior is prepared in two forms: Saffman type $\mathcal R_C\subset \mathcal R_B$8 and Batchelor type $\mathcal R_C\subset \mathcal R_B$9 (Yu et al., 2020). The paper ties the low-$|s|<\mu$0 coefficients to the Saffman integral $|s|<\mu$1 and the Loitsyansky integral $|s|<\mu$2, interpreting nonzero impulse as leading to $|s|<\mu$3 and zero impulse as leading to $|s|<\mu$4 (Yu et al., 2020). A compensating Stokes vortex ring is added in the $|s|<\mu$5 case to cancel the initial impulse created by windowing and projection (Yu et al., 2020).

The dynamics show a decaying but expanding turbulent cloud. The mean radius is defined by

$|s|<\mu$6

with $|s|<\mu$7, and both $|s|<\mu$8 and the integral scale show similar power-law-like growth (Yu et al., 2020). For energy decay and integral-scale growth, the classical predictions are

$|s|<\mu$9

for Saffman $$\Phi(r)=\frac{1}{2}\left[1-\tanh\left(\frac{2(r-R)}{\sigma}\right)\right]$$0 turbulence and

$$\Phi(r)=\frac{1}{2}\left[1-\tanh\left(\frac{2(r-R)}{\sigma}\right)\right]$$1

for Batchelor $$\Phi(r)=\frac{1}{2}\left[1-\tanh\left(\frac{2(r-R)}{\sigma}\right)\right]$$2 turbulence (Yu et al., 2020). The observed decay in the spherical turbulent cloud is closer to the Saffman prediction in both cases, at least up to about 400 initial eddy turnover times, and the low-wavenumber distinction has little effect on the inertial scales over that time span (Yu et al., 2020). The paper further reports that the coefficient of the $$\Phi(r)=\frac{1}{2}\left[1-\tanh\left(\frac{2(r-R)}{\sigma}\right)\right]$$3 term evolves in time in the Batchelor-type case, indicating that the Loitsyansky integral is not constant in this finite, inhomogeneous flow (Yu et al., 2020).

The most distinctive SRoI-edge phenomenon is the late-time ejection of vortex rings from the cloud boundary (Yu et al., 2020). These rings have a size comparable to the initial integral scale; changing the ratio $$\Phi(r)=\frac{1}{2}\left[1-\tanh\left(\frac{2(r-R)}{\sigma}\right)\right]$$4 changes the ejection size, whereas changing $$\Phi(r)=\frac{1}{2}\left[1-\tanh\left(\frac{2(r-R)}{\sigma}\right)\right]$$5 or $$\Phi(r)=\frac{1}{2}\left[1-\tanh\left(\frac{2(r-R)}{\sigma}\right)\right]$$6 has little effect (Yu et al., 2020). The proposed mechanism is a local imbalance of impulse, measured by the Gaussian-weighted quantity

$$\Phi(r)=\frac{1}{2}\left[1-\tanh\left(\frac{2(r-R)}{\sigma}\right)\right]$$7

whose maximum peaks when $$\Phi(r)=\frac{1}{2}\left[1-\tanh\left(\frac{2(r-R)}{\sigma}\right)\right]$$8 is on the order of the integral scale (Yu et al., 2020). Near the cloud edge, outward-pointing local impulse imbalance can eject vorticity into the ambient fluid and generate vortex rings (Yu et al., 2020). In SRoI terms, this is an example where the spherical boundary is not only a geometric delimiter but also a dynamically active interface.

5. Region-wide control and representation on spherical domains

In acoustics, the principal issue is not truncation but region-wide reproduction of perceptually relevant field quantities. The paper on reproducing acoustic velocity vectors in a spherical listening region introduces velocity matching (VM), which reproduces AVVs by matching their spherical harmonic coefficients throughout the entire spherical region rather than at discrete sweet spots (Wang et al., 2023). Pressure measured by a higher-order microphone array is translated into local coefficients and then into AVV coefficients; because the velocity at a point depends only on the $$\Phi(r)=\frac{1}{2}\left[1-\tanh\left(\frac{2(r-R)}{\sigma}\right)\right]$$9 local SH terms, the AVV field admits a compact representation in terms of the pressure SH coefficients (Wang et al., 2023). The reproduction equation is

$S^2$0

contrasted with pressure matching

$S^2$1

with weights obtained using the Moore-Penrose pseudoinverse (Wang et al., 2023). In a simulation with an 8-channel loudspeaker array on a sphere of radius $S^2$2 m and a listening region of radius $S^2$3 m, the error was averaged over 113081 evaluation points in the larger sphere and 2993 evaluation points in a central $S^2$4 m sphere; VM gave lower AVV reproduction error than pressure matching below about 500 Hz in the $S^2$5 m sphere and up to slightly above 1 kHz in the smaller sphere (Wang et al., 2023). The paper also states that VM requires fewer loudspeakers than pressure matching for accurate low-frequency AVV reproduction (Wang et al., 2023).

A different spherical-region construction appears in 360° image super-resolution. The SphereSR framework defines a continuous image representation on the sphere,

$S^2$6

and reconstructs RGB values at arbitrary spherical coordinates using the Spherical Local Implicit Image Function (SLIIF) (Yoon et al., 2021). For a query point inside an icosahedral face, the decoded RGB value is

$S^2$7

with $S^2$8 the sinusoidal positional encoding (Yoon et al., 2021). The underlying geometry is local and sphere-aware: each query is reconstructed from its containing face, three surrounding vertices, and local vertex neighborhoods on the subdivided icosahedron (Yoon et al., 2021). The paper does not define an SRoI explicitly, but its face-based local spherical decoding and cell-aware reconstruction of projected pixel areas are closely aligned with SRoI-like processing on $S^2$9.

Wide-field radio-interferometric imaging provides another spherical-domain analogue. S-R2D2 extends the planar R2D2 paradigm to the sphere by factorizing the wide-field measurement operator into a sphere-to-plane interpolator $C\subset B$00 and the standard planar RI operator $C\subset B$01: $C\subset B$02 The reconstruction is physically interpreted on the sphere through $C\subset B$03, while the DNN remains a 2D U-Net acting on planar tensors (Tajja et al., 3 Mar 2025). The field of view is a spherical cap with $C\subset B$04, sampled on HEALPix with $C\subset B$05 and $C\subset B$06 spherical pixels within the FOV (Tajja et al., 3 Mar 2025). The ideal planar resolution rule would require $C\subset B$07, but the paper uses $C\subset B$08, $C\subset B$09, and $C\subset B$10 to preserve scalability, explicitly learning to correct the resulting interpolation approximations through a spherical-domain loss (Tajja et al., 3 Mar 2025). Averaged test results at final iteration $C\subset B$11 reported S-R2D2 SNR/logSNR of $C\subset B$12 dB at $C\subset B$13, $C\subset B$14 dB at $C\subset B$15, and $C\subset B$16 dB at $C\subset B$17, compared with markedly lower values for planar R2D2 (Tajja et al., 3 Mar 2025). Here the spherical region is global rather than local, but the governing principle is similar: the geometry of $C\subset B$18 cannot be treated as an incidental discretization choice.

6. Common themes, distinctions, and limitations

The literature does not present a single universally standardized meaning of SRoI across fields. In CT, the term is explicit and denotes a spherical subvolume $C\subset B$19 selected for truncated-data reconstruction (Azencott et al., 2015). In turbulence, the phrase is an interpretive extension applied to a compact spherical turbulent cloud embedded in free space (Yu et al., 2020). In acoustics, the relevant object is a spherical listening region covering a full volume rather than a boundary or a sparse set of points (Wang et al., 2023). In spherical image and radio-interferometric work, the closest analogues are local spherical neighborhoods or spherical fields of view rather than a named SRoI (Yoon et al., 2021, Tajja et al., 3 Mar 2025). A plausible implication is that SRoI functions as a cross-domain organizing concept rather than a domain-invariant formalism.

Despite these differences, three recurring structural themes emerge. First, spherical geometry is used to impose rotationally symmetric localization, whether for acquisition, initialization, or control. Second, the sphere generally interacts with global operators rather than replacing them: CT truncation is defined through ray sets, turbulence requires exact free-space physics outside the sphere, and acoustics represents the field throughout the entire spherical volume. Third, spherical localization typically exposes boundary-sensitive phenomena. In tomography this appears as missing-data instability and null-space artifacts (Azencott et al., 2015, Han et al., 2017); in turbulence it appears as TNTI corrugation and vortex-ring ejection (Yu et al., 2020).

Common misconceptions follow from conflating these roles. An SRoI is not always a crop from a larger image, because in CT the missing rays outside the ROI directly change the inverse problem (Azencott et al., 2015). It is not always a closed isolated system, because the surrounding medium may remain dynamically essential, as in free-space turbulence (Yu et al., 2020). Nor does spherical localization automatically imply universal large-scale behavior: the turbulence study explicitly reports that the largest scales are controlled by the initial condition and by the finite-domain construction, especially at very low wavenumber (Yu et al., 2020).

For future work, the existing papers suggest several directions without resolving them uniformly. One is the extent to which spherical localization should be paired with exact geometry-aware operators rather than Euclidean approximations. Another is whether learned correctors, such as CNN null-space suppressors in interior tomography or interpolation-aware training in spherical RI, can be given guarantees comparable to contraction-based iterative inverses in ROI CT (Han et al., 2017, Tajja et al., 3 Mar 2025, Azencott et al., 2015). A further implication is that SRoI-based methods may be most reliable when they separate interior fidelity, interface behavior, and exterior consistency rather than assuming that the sphere alone regularizes the problem.