Organizations: The University of Hong Kong · The University of Sydney · Chongqing University · University of Science and Technology of China · Shenzhen Bay Laboratory
Sparse-view and limited-arc computed tomography (CT) are commonly treated as projection completion: a network fills an unmeasured view, and the result is ranked by similarity to one ground-truth image. That score can be high even when the query angle is far from every acquired measurement, because a learned prior may invent anatomy the scan never constrained. Existing protocols therefore leave a basic question unanswered: how does predictive reliability change with distance from the acquired views? We propose the Angular Calibrated Uncertainty Score (ACUS) to answer that question. ACUS is the Winkler interval score of a nominal 90% predictive interval on held-out projections, binned by distance to the nearest acquired view. It charges interval width and adds a large miss penalty when a held-out pixel falls outside the interval, so a narrow band helps only when coverage holds. We use collapse on acquired angles as a necessary consistency check on the predictor. On chest and dental cone-beam orbits, interval width grows with the angular gap. Empirical coverage ranges from 77% to 92%, and the peak signal-to-noise ratio (PSNR) of the point estimate falls as the width rises. Nearest-neighbor, interpolation, and point completion, each given one residual scale, have zero reported far coverage and incur outside-distance penalties. ACUS exposes gap-dependent overconfidence that a point-estimate PSNR alone does not measure.
Figures & tables
Figure 1: Incomplete-view CT as an interval problem. (a) Sparse-view and limited-arc scans leave angles with no ray coverage. A prior can complete them, and the question is whether the image is trustworthy. (b) PSNR, SSIM, and acquired-view geometry give no error bound on those angles. (c) ACUS reports width and coverage by the gap to the nearest acquired view: narrow nearby, wide far away. A static scale fails at large gaps.
Figure 2: Interval estimator used to compute ACUS. (a) A frozen MAE encodes acquired views. Distance-weighted softmax fuses the two nearest into a mean μ and a scale σ . Point completion is the mean-only baseline. δ marks acquired, near, and far queries. (b) Chest and dental specialists train on sparse and multi-arc orbits with dropout and held-out queries. Gaussian likelihood has no variance target. Collapse drives σ toward zero and μ toward the acquired view. (c) The nominal 90% interval is μ±1.645σ (width 3.29σ ). Inside, ACUS equals the width. Outside, it adds twenty times the distance to the bound.
Method
u near/far
Cover near/far
PSNR near/far
Chest
Nearest
0.014/0.014
0.61/0.00
30.27/16.40
Linear
0.012/0.012
0.70/0.00
32.34/16.34
Point ( ℓ1 )
0.013/0.013
0.72/0.00
32.17/14.14
ACUS
0.018/0.088
0.81/0.87
32.48/18.64
Dental
Table 1: Sparse K=8 , 40 test cases per domain. Entries are near/far. u is mean scale, not interval length. Cover targets 0.90 . Point denotes point completion. ACUS scale standard deviations (near/far): Chest 0.003/0.003 , Dental 0.002/0.002 .
Method
u near/far
Cover near/far
PSNR near/far
Chest
Nearest
0.022/0.022
0.74/0.00
29.05/15.51
Linear
0.022/0.022
0.71/0.00
29.18/15.28
Point ( ℓ1 )
0.026/0.026
0.77/0.00
28.79/13.59
ACUS
0.023/0.095
0.77/0.80
29.59/17.48
Dental
Table 2: Multi-arc K=8 , 40 test cases per domain. Entries are near/far. u is mean scale, not interval length. Cover targets 0.90 . Point denotes point completion. ACUS scale standard deviations (near/far): Chest 0.003/0.008 , Dental 0.001/0.002 . Acquired scale is 0.002 in both domains (PSNR 52.6/54.9 dB).
Figure 3: Far-angle held-out projections on chest and dental cone-beam orbits. Columns: ground truth (GT), nearest neighbor (NN), point completion, ACUS mean, ACUS scale, color residual maps for NN, point completion, and ACUS, and miss masks, with a shared residual colorbar. Constant-scale baselines miss most far pixels. ACUS keeps a structured scale and a sparse miss mask.
Input: acquired set A with views yA , query angle a′ with held-out projection y , predictor (μ,σ) , level α=0.1 .
Output: ACUS(a′) and mean scale u(a′) .
1: δ(a′)←mina∈A∣a′−a∣
2: (μ,σ)←Predict(a′;yA) using the two nearest views
3: L←μ−1.645σ , U←μ+1.645σ
4: for each detector pixel i do
5: Si←Ui−Li . If yi∈/[Li,Ui] then Si←Si+α2di
Table 3: Computing ACUS for one query angle.
Figure 4: Same far-query protocol with classical and learned reconstructors. (a) Chest. (b) Dental. Each method shows prediction, color residual, and miss mask with a shared colorbar. Feldkamp-Davis-Kress (FDK) [ 10 ] and Simultaneous Algebraic Reconstruction Technique (SART) [ 1 ] reconstruct the observed orbit and are reprojected at the query angle. Structure-Aware X-ray Neural Radiodensity Fields (SAX-NeRF), R 2 -Gaussian, Neural Attenuation Fields (NAF), and X-Gaussian use the same orbit geometry [ 7 , 25 , 26 , 6 ] .
Figure 5: Acquisition orbit schematics for chest and dental sparse protocols. Colored segments mark unobserved views on the 50-view, 180∘ orbit, using the ACUS color scale.
Figure 6: Orbit rings with ACUS-colored borders and a central volume rendering. (a) Chest. (b) Dental. Cold borders sit near acquired views. Warm borders sit on far queries. The central volume is a spatial anchor, not a reconstruction in the measurement fiber.
Figure 7: Acquired-view collapse and cross-anatomy far coverage on multi-arc queries. (a) On the 20-case multi-arc ablation split, distinct from the 40-case main tables, removing Lcol raises the acquired scale from 0.002 to 0.018 (chest) and 0.015 (dental). Far-view scales remain 0.095 (chest) and 0.071 (dental). (b) Far coverage for official 40-case same-domain cells and the official transfer evaluation. The dental-to-chest collapse is a negative control, not a transfer success.
Chest, sparse
Dental, sparse
Gap ( ∘ )
u
Cover
PSNR
u
Cover
PSNR
0 to 10
0.014±0.001
0.791±0.069
33.88±1.56
0.021±0.001
0.931±0.024
33.17±1.04
10 to 20
0.027±0.005
0.848±0.063
29.49±2.47
0.030±0.002
0.911±0.043
29.15±1.83
20 to 30
0.055±0.009
0.851±0.066
23.11±1.71
0.047±0.004
0.900±0.047
24.71±1.98
30 to 40
0.079±0.004
0.833±0.087
19.18±1.50
0.057±0.001
0.860±0.051
22.00±0.57
40 to 60
0.088±0.003
0.869±0.052
18.64±0.99
0.061±0.002
0.858±0.083
21.12±2.04
Table 4: Official 40-case K=8 angular bins for ACUS. u is mean scale, not interval length. Cover targets 0.90 . PSNR is of the predictive mean, in dB. Sparse protocols have no 60∘ to 180∘ bin. Multi-arc far columns in Tab. 2 pool the 40∘ to 60∘ and 60∘ to 180∘ bins. The last row has limited case support.
Figure 8: Sparse K=8 gap bins on the official 40-case split. (a) Chest. (b) Dental. Curves show mean scale, empirical coverage, and PSNR of the predictive mean. The 40∘ to 60∘ bin is the far column of Tab. 1 . Error bars are case-level standard deviations, and the dashed line is the nominal 90% rate. Mean ACUS is omitted because the outside-distance term is unreported.
Figure 9: Official 643 slices side by side in one column, chest on the left and dental on the right. (a) Chest. (b) Dental. Columns are axial, coronal, and sagittal. Rows are the ground-truth volume, sparse FDK, and a free residual refinement that can look closer to the anatomy while agreeing less with the acquired projections, and does not yield a volume in F ( Section 4 ).
Appendix figures & tables33 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 10: Official K=8 bins, 40 cases per domain. Bands are case-level standard deviations. The dashed line is the nominal 90% rate. The 60∘ to 180∘ chest and dental multi-arc points have no case-level deviation because that bin contains one query pattern in the official split. These curves use the printed supplement tables, not a later rerun.
Figure 11: Held-out 32.4∘ mid-range intervals in the 30∘ to 40∘ bin, not a far query of at least 40∘ . Panels show ground truth (GT), linear interpolation, the ACUS mean, a scale map, the 90% bounds, the residual, and outside-interval pixels.
Figure 12: Acquired-view collapse on the 20-case multi-arc ablation split. The panel is the same acquired-view collapse diagnostic as main Fig. 7 (a). The values agree with Tab. 10 .
Figure 13: Near versus far readout on one chest sparse case. Left block: near query. Right block: far query. Color panels show scale and absolute residual with a shared colorbar. Near the acquired view every method looks similar. Far away only ACUS reports a non-constant scale and keeps most pixels inside the interval.
Figure 14: Dental near versus far with color scale and residual maps and a shared colorbar. The dental specialist keeps a narrower absolute scale than the chest specialist, while far coverage remains near the nominal rate on its own domain.
Figure 15: Chest multi-arc near versus far with color residuals. Multi-arc packing changes which angles are local, but the same near-narrow and far-wide pattern appears.
Figure 16: External learned methods with color residuals on chest and dental far queries. Each row is one reconstructor. Columns alternate prediction, residual, and miss for chest and dental.
Figure 17: External methods with residual zoom insets on chest far queries. Colored boxes mark local regions. The bottom-right zooms enlarge residual structure that a full-frame thumbnail hides.
Figure 18: Residual and miss mosaic across nearest neighbor (NN), point completion, and ACUS under four protocols. Warm residual mass on far queries is the qualitative signature of an under-wide interval.
Figure 19: Chest sparse overview with local zoom insets. Top row: full panels with colored region-of-interest (ROI) boxes. Bottom row: magnified crops of ACUS mean, scale, and residual.
Figure 20: Dental multi-arc overview with local zoom insets. The dental scale is smaller in absolute value, and the residual zooms remain structured and locally organized.
Figure 21: Chest multi-arc overview with local zoom insets. Comparing Figs. 19 , 20 and 21 shows that the zoom pattern tracks the gap, not the packing of the acquired arcs.
Figure 22: Protocol wall across eight settings. Rows: ground truth, ACUS mean, ACUS scale, absolute residual. Columns: sparse, multi-arc, near, and mixed near protocols on chest and dental.
Figure 23: ACUS scale heatmap wall across six protocols. The scale opens on far settings and stays dark near acquired views.
Figure 24: Cross-anatomy and detailed qualitative panels. (a) Cross-anatomy transfer. (b) to (d) Domain-specific detail boards.
Figure 25: Additional chest qualitative panels from held-out orbits. (a) to (e) are distinct cases, each showing a query strip, the held-out ground-truth projection, and the predicted projection. These panels are included to show breadth across cases, not to claim a new metric.
Figure 26: Combined 3D orbit, volume-slice, collapse, and cross-anatomy board. The volume slice row is a negative diagnostic from the main text. Higher volume PSNR can coexist with worse agreement on acquired projections.
Figure 27: Volume-slice diagnostic and acquired-view collapse maps. (a) and (b) Axial and related slices from the 643 free residual refinement. (c) and (d) Acquired-view scale maps that should collapse near zero.
Figure 28: Width-versus-angle curve and reprojection diagnostic for the volume readout experiment.
Figure 29: Collapse gate on acquired views for chest and dental specialists.
Figure 30: Sparse far-query detail boards for chest and dental.
Figure 31: Multi-arc far-query detail boards for chest and dental.
Gap
u
Cover
PSNR
0∘ to 10∘
0.014±0.001
0.791±0.069
33.88±1.56
10∘ to 20∘
0.027±0.005
0.848±0.063
29.49±2.47
20∘ to 30∘
0.055±0.009
0.851±0.066
23.11±1.71
30∘ to 40∘
0.079±0.004
0.833±0.087
19.18±1.50
40∘ to 60∘
0.088±0.003
0.869±0.052
18.64±0.99
Appendix
Table 5: Chest sparse bins (40 cases), from the same official K=8 split as the main sparse table. u is mean scale, not interval length. The 40∘ to 60∘ row is the far column of that table.
Gap
u
Cover
PSNR
0∘ to 10∘
0.013±0.002
0.750±0.062
32.88±1.30
10∘ to 20∘
0.034±0.005
0.785±0.066
25.75±1.65
20∘ to 30∘
0.057±0.008
0.859±0.053
23.15±1.55
30∘ to 40∘
0.075±0.012
0.868±0.062
21.10±2.31
40∘ to 60∘
0.094±0.008
0.800±0.054
17.54±0.53
60∘ to 180∘
0.111±0.000
0.804±0.000
16.68±0.00
Appendix
Table 6: Chest multi-arc bins (40 cases), from the same official split as the main multi-arc table. u is mean scale, not interval length. Far queries aggregate the 40∘ to 60∘ and 60∘ to 180∘ bins.
Gap
u
Cover
PSNR
0∘ to 10∘
0.021±0.001
0.931±0.024
33.17±1.04
10∘ to 20∘
0.030±0.002
0.911±0.043
29.15±1.83
20∘ to 30∘
0.047±0.004
0.900±0.047
24.71±1.98
30∘ to 40∘
0.057±0.001
0.860±0.051
22.00±0.57
40∘ to 60∘
0.061±0.002
0.858±0.083
21.12±2.04
Appendix
Table 7: Dental sparse bins (40 cases), from the same official split as the main sparse table. u is mean scale, not interval length. The 40∘ to 60∘ row is the far column of that table.
Gap
u
Cover
PSNR
0∘ to 10∘
0.021±0.001
0.914±0.029
32.70±1.27
10∘ to 20∘
0.034±0.002
0.853±0.055
26.38±1.40
20∘ to 30∘
0.047±0.003
0.862±0.064
24.16±1.51
30∘ to 40∘
0.055±0.003
0.870±0.067
22.85±1.68
40∘ to 60∘
0.062±0.002
0.907±0.025
22.41±0.81
60∘ to 180∘
0.068±0.000
0.914±0.000
21.59±0.00
Appendix
Table 8: Dental multi-arc bins (40 cases), from the same official split as the main multi-arc table. u is mean scale, not interval length. Far queries aggregate the 40∘ to 60∘ and 60∘ to 180∘ bins.
Table 10: Collapse ablation on a 20-case multi-arc split.
Train → Test
Far width
Far cover
Far PSNR
Chest → Dental
0.093±0.013
0.84
19.59±2.08
Dental → Chest
0.061±0.001
0.59
16.58±1.12
Appendix
Table 11: Cross-anatomy transfer on multi-arc orbits.
Domain
Near width
Far width
PSNR near/far
Far cover
Chest
0.027
0.090
32.45/18.68
0.82
Dental
0.027
0.064
31.25/21.28
0.88
Appendix
Table 12: Long-arc stress test (16 cases, one arc of 50∘ to 120∘ ).
Domain
Method
Vol. PSNR
Reproj. PSNR
Query interval inclusion
Chest
sparse FDK
6.82±0.36
13.32±1.41
n/a
Chest
free residual refinement
14.04±1.11
9.79±1.10
0.07 to 0.15
Dental
sparse FDK
9.41±0.56
10.41±0.99
n/a
Dental
free residual refinement
18.26±0.82
9.13±0.87
0.07 to 0.15
Appendix
Table 13: Volume readout diagnostic on 12 sparse cases at 643 .
Domain
Orbit
Acq. width
Pred. PSNR on acquired
Chest
Sparse
0.002
46.4
Chest
Multi-arc
0.002
52.6
Dental
Sparse
0.002
49.5
Dental
Multi-arc
0.002
54.9
Appendix
Table 14: Acquired-view collapse check for ACUS (main K=8 cells).
Figure 32: Chest multi-arc far coverage on the later rerun, N=40 , K=8 . Gray bars use one residual scale for every angle. Blue bars use one residual scale per gap bin. The line is the learned scale from that same rerun, not the official multi-arc cell.
Radiative Gaussian splatting reconstructs sparse-view CT fast and accurately, and recent work attaches per-Gaussian posteriors to yield per-voxel uncertainty maps. We ask what such a map actually measures: posterior variance is a data-constraint map, not an error map -- its alarms are trustworthy, its all-clears are not. Exploiting the strict linearity of X-ray rendering in the per-Gaussian densities, we derive a clamp-aware closed form that the unchanged rasterizer evaluates exactly in one forward pass, in volume and projection space: the infinite-sample limit of the sampling estimator of concurrent work, at ~8x lower cost. On the official 15-scene benchmark this uncertainty ranks true error on 14 of 15 scenes. Restricted to the object interior -- the tissue a clinician reads -- the ranking collapses (median Spearman 0.11, 0/15 pass), identically for a deep ensemble and for a strictly positive log-normal posterior: three constructions, two estimator families, no survivors. The mechanism is structural: about 90% of in-object error is bias that reproduces across retrainings, invisible to model disagreement; 73-81% of the full-volume correlation is carried by object/surround contrast; and an exactly solvable control puts the observed in-object ranking 4-5x below what a perfectly calibrated posterior with the same sigma-spread would score. The error scale, by contrast, is an engineering problem, and we solve it: reparameterizing the posterior contracts the cross-scene temperature spread from 19.3x to 2.6x, one scene-agnostic temperature transfers to unseen scenes (10/15 leave-one-scene-out), and the repaired scale tracks photon count at the Poisson-predicted -1/2 power. We distill evaluation practice that would have caught the illusion -- masked calibration, seed-wise bias decomposition, an exact-posterior reference -- and release all protocols, seeds and per-run evidence.
Chulin Zhao, Yiran Xu, Shu Liu
1Dundee International Institute, Central South University · 2Central South University
Sparse-view computed tomography (CT) reduces radiation dose and acquisition time and may mitigate motion artifacts. However, angular undersampling provides insufficient information to determine the image uniquely and stably. Limited-angle CT, arising from restricted angular coverage, produces strongly directional information loss associated with the missing angular range. In both settings, image directions may be strongly observed, weakly constrained, or unobservable, leading to severe ill-posedness and reconstruction ambiguity. Existing diffusion-based approaches incorporate measurement information through likelihood guidance, data-consistency operations, or range-null-space corrections. However, they do not generally use the continuously varying measurement sensitivity of the acquisition to jointly shape both reconstruction updates and stochastic exploration. We propose a geometry-aware diffusion-guided stochastic reconstruction framework for sparse-view and limited-angle CT. Its central component is a regularized noise-weighted pullback metric constructed from the CT forward operator and measurement-noise covariance. This metric continuously adapts both the measurement-aware update and stochastic exploration according to directional measurement sensitivity, suppressing changes along strongly constrained directions while permitting greater exploration along weakly constrained and unobservable directions. We complement this geometry-aware update with a regularized data-consistency correction and approximately null-space-restricted stochastic perturbations, implemented matrix-free using forward and backprojection operations together with conjugate-gradient solves. Experiments on sparse-view, noisy, and limited-angle CT demonstrate competitive reconstruction quality, strong measurement consistency, and spatially resolved empirical uncertainty estimates.
Honglei Brinkmann, Carola-Bibiane Schönlieb, Ander Biguri
Department of Applied Mathematics and Theoretical Physics, University of Cambridge, Cambridge, UK.
Sparse-view computed tomography (CT) reduces radiation dose and acquisition time but remains severely ill-posed because incomplete projections poorly constrain null-space information. Existing learning-based methods often estimate this information in high-dimensional image space, conflate physical measurement errors with prediction errors, and depend on fixed discretizations. We propose SONAR, a Structure-Consistent Neural Operator for Null-Space-Aware Reconstruction. Instead of recovering the full null-space component, SONAR predicts a low-dimensional null-space-aware representation from the acquired projections as pseudo-measurements. It separates measurement and pseudo-measurement residuals, lifts them into the image domain through physics operators, and applies independent neural operators to constrain their structural effects, thereby accommodating admissible errors while suppressing unsupported structures. To support cross-discretization reconstruction, an anisotropic U-shaped neural operator models the periodic angular and nonperiodic detector dimensions using direction-dependent continuous supports, while image-domain neural operators re-discretize continuous kernels on target grids. These components form an optimization-inspired unrolled network. Experiments on simulated AAPM and clinical MARS photon-counting CT data demonstrate consistent improvements across seen and unseen view settings and unseen image resolutions. On AAPM dataset, SONAR improves PSNR by 1.87dB at 62 views and by 7.63dB under zero-shot transfer to a 512×512 grid over the strongest competing methods. SONAR also achieves the best overall performance in all clinical settings evaluated, demonstrating accurate, structurally reliable, and discretization-robust sparse-view CT reconstruction.
Song Ni, Haijun Yu, Haodong Li +4
Department of Electrical and Computer Engineering, University of Massachusetts Lowell, Lowell, MA 01854