Beam Geometry and Input Dimensionality: Impact on Sparse-Sampling Artifact Correction for Clinical CT with U-Nets

Daniela Pfeiffer; Florian Schaff; Franz Pfeiffer; Johannes Thalhammer; Sebastian Peterhansl; Tina Dorosti

REVIEW 4 major objections 4 minor 1 references

Beam Geometry and Input Dimensionality: Impact on Sparse-Sampling Artifact Correction for Clinical CT with U-Nets

T0 review · 4 major / 4 minor · reviewed 2026-08-15 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read The paper claims that across parallel, fan, and cone beam CT geometries, a 2D U-Net trained on full $512\times512$ axial slices surpasses 2.5D and 3D U-Nets that use $64^3$-voxel block inputs for sparse-sampling streak artifact correction.

desk verdict The 2D advantage is plausible but the comparison confounds dimensionality with resolution and field of view; the geometry sweep is still a useful benchmark. read the letter →

arxiv 2508.17961 v1 pith:VOH4YK65 submitted 2025-08-25 cs.CV physics.med-ph

Tina Dorosti , Johannes Thalhammer , Sebastian Peterhansl , Daniela Pfeiffer , Franz Pfeiffer , Florian Schaff This is my paper · ORCID

classification cs.CVphysics.med-ph

keywords sparse-samplingCTstreakartifactcorrectionU-Netinputdimensionality2.5D3Dbeamgeometryclinical

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper sets out to test whether adding volumetric context to a U-Net improves correction of sparse-sampling streak artifacts in clinical CT, and whether any such benefit depends on beam geometry. Using 22 retrospectively selected patient scans, the authors simulate sparsely sampled reconstructions in parallel, fan, and cone beam geometries, then train 2D, 2.5D, and 3D U-Nets. They find that the 2D U-Net trained on full $512\times512$ axial slices gives the best MSE and SSIM for all three geometries, beating both the 2.5D representation (three orthogonal $64\times64$ center-plane patches arranged as a $64\times64\times3$ image) and the 3D representation ($64^3$-voxel blocks). The result matters because sparse-sampling CT lowers dose, and knowing that a simple full-slice 2D network outperforms volumetric block inputs could simplify model choice.

What carries the argument

The load-bearing machinery is the contrast among three input representations fed to the same U-Net family, a convolutional encoder-decoder architecture with skip connections. The 2D condition uses the full $512\times512$ axial CT slice. The 3D condition divides each CT volume into $64\times64\times64$ voxel blocks and treats each block as a volume input. The 2.5D condition takes the axial, coronal, and sagittal cut through the center of each block, stacks the three $64\times64$ patches as channels of a single $64\times64\times3$ image, and feeds that to a 2D network. Streak artifacts are generated by sparse-view simulation from full-view 2048-projection scans. The argument turns on the comparison across these representations: because MSE and SSIM are evaluated on the same eight test subjects, any consistent ranking across geometries is attributed to input dimensionality and the resolution at which each network sees the artifact.

What would settle it

Retrain the 2.5D and 3D networks using full-resolution $512\times512\times n$ volumetric patches, or 2.5D inputs assembled from full $512\times512$ center-plane slices, on the same fourteen training subjects and evaluate on the same eight test subjects. If the 2D model still has the best MSE and SSIM, input dimensionality is the deciding factor; if block-based models catch up, the 64-voxel block cropping was the confound and the paper's stated conclusion would need revision.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is a negative one about dimensionality: for sparse-sampling streak artifact correction in clinical CT, giving a U-Net volumetric context through 2.5D or 3D inputs does not help and actually hurts. For every beam geometry tested (parallel, fan, and cone beam), the 2D U-Net operating on full $512\times512$ axial slices attains the best mean squared error and structural similarity index, outperforming the 2.5D input formed by stacking three $64\times64$ orthogonal center-plane patches into a $64\times64\times3$ image and the 3D input formed by $64^3$-voxel blocks. The authors interpret this as evidence that full-resolution axial slices already carry the information needed to remove streak artifacts, and that cropping into blocks to make room for volumetric context is counterproductive. The ranking is consistent across geometries, so the dimensionality finding is not tied to one acquisition setup.

Load-bearing premise

The load-bearing assumption is that differences between the input conditions are due to dimensionality alone, but the 2D network sees a full $512\times512$ slice while the 2.5D and 3D networks see only $64^2$- or $64^3$-voxel blocks, so resolution and field of view change at the same time as dimensionality.

Editorial extensions

If this is right

A 2D U-Net on full axial slices is the better default for sparse-sampling streak correction on these clinical CT data, regardless of whether the acquisition is parallel, fan, or cone beam.
Switching to 2.5D or 3D inputs built from 64-voxel blocks does not buy accuracy; it costs MSE and SSIM relative to the full-slice 2D baseline.
The 2.5D trick of stacking three orthogonal center-plane patches does not recover the context lost in block cropping.
Full-view 2048-projection simulations provide a consistent reference, so the dimensionality ranking is measured against the same ground truth across geometries.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The 2.5D and 3D conditions are built from $64\times64\times64$ blocks, while the winning 2D condition uses the full $512\times512$ slice, so the reported comparison does not isolate dimensionality from resolution and field of view; a 3D model given full-resolution volumetric context might reasonably be expected to do better than this experiment shows.
An untested extension is to keep the in-plane resolution at $512\times512$ and add depth, for example $512\times512\times n$ patches or multi-slice 2.5D inputs; if the 2D model still wins, dimensionality itself is the cause, and if not, block cropping was the cause.
Streak artifacts from sparse sampling are global structures spanning the whole image, so cropping into 64-voxel blocks may remove the very context the network needs; this suggests context, not dimensionality, is the operative variable.
The consistent result across beam geometries is compatible with the view that these streaks are largely correctable slice-by-slice on these data, but a direct test would require ablating in-plane resolution while holding dimensionality fixed.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Quick take: the headline result—2D beat 2.5D and 3D across all three beam geometries—is a useful practical data point, but the comparison as described doesn't isolate dimensionality. The 2D condition uses full 512x512 slices; the 2.5D and 3D conditions are built from 64x64x64 blocks, with 2.5D keeping only three center-plane cuts. So the 2D network sees the whole slice and the entire streak pattern, while the volumetric networks only see a small block. That alone could explain the reported advantage. The stress-test note lands.

What's genuinely new and worth credit: the systematic sweep over parallel, fan, and cone beam geometries on 22 retrospective clinical CT volumes, with a 2048-view reference and a fixed U-Net architecture. That's a clean benchmark frame for a practical engineering question. If the result holds, it would say: don't bother with small-block volumetric context for this artifact-correction task; full-resolution 2D axial is the default. That's valuable to people building sparse-view CT pipelines.

The soft spots, in proportion. The confound is load-bearing for the claim; without a resolution-matched control, the abstract overreaches. The abstract also reports no error bars or significance tests, and with 8 test subjects the margins matter. Those are fixable reporting gaps, not fatal problems.

I could only assess the abstract: the supplied full-text extraction is corrupted and interleaved with an unrelated federated-learning preprint, so I can't verify whether the paper already includes a matched-resolution control or a per-subject breakdown. The UNVERDICTED status is the right call under that constraint.

Bottom line: this is a plausible, potentially useful empirical study with a central claim that needs one more control. A serious referee should look at the full paper, mainly to check whether the missing comparison exists in the body and whether the effect sizes are large enough to overcome the block-resolution disadvantage. If not, the conclusion should be softened to "2D axial is a stronger default under our preprocessing" rather than "2D beats volumetric inputs because dimensionality."

Recommendation: send to peer review—the question is practical, the data is real, and the confound is checkable. Expect major revision unless the full text already addresses it.

Referee Report

4 major / 4 minor

Summary. The manuscript reports an empirical study of U-Net-based sparse-sampling streak artifact correction in clinical CT, comparing 2D, 2.5D, and 3D input representations across parallel, fan, and cone beam geometries. Using 22 retrospectively selected subjects (14 for training/validation and 8 for testing) and simulated sparse projections via the Astra toolbox, the authors report MSE and SSIM and conclude that the 2D axial-slice U-Net achieves the best values for all geometries. The abstract's preprocessing description states that 2D inputs are full 512x512 axial slices, while 2.5D and 3D inputs are built from 64x64x64 voxel blocks; 2.5D keeps only three 64x64 center-plane cuts arranged as a 64x64x3 image, and 3D uses individual 64-voxel blocks.

Significance. If the comparison were clean, the result would be practically relevant: it would indicate that volumetric context, at least in the specific block-based forms tested, does not improve streak artifact correction and that a simpler 2D network may be the preferred default across beam geometries. The study addresses a clinically motivated question, uses real patient data, and covers three beam geometries, which are strengths. However, the load-bearing comparison is confounded, and no uncertainty quantification is reported, so the stated conclusion is not currently supported. The manuscript does not ship code, trained models, or machine-checked proofs; its contribution is an empirical benchmark that needs a cleaner experimental design before the central claim can be accepted.

major comments (4)

[Abstract (pre-processing description)] The comparison presented as isolating input dimensionality does not isolate it. The 2D condition trains on full 512x512 axial slices, while the 2.5D and 3D conditions use 64x64x64 voxel blocks; 2.5D keeps only three 64x64 center-plane cuts and 3D uses individual 64-voxel blocks. Resolution, field of view, and the amount of global streak context available to the network therefore change together with dimensionality. Sparse-sampling streak artifacts are global across the field of view, so a 64x64 block cannot represent them, while a 512x512 slice can. The reported 2D advantage is fully explainable by information loss in the block representation rather than by dimensionality. The central claim requires either an experiment that matches resolution and field of view across conditions (for example, 2D on equivalent 64x64 patches, or volumetric models on full-FOV inputs) or a substantially narrowed claim that does not attribute the result to input dimensionality.
[Abstract (results)] The abstract reports only qualitative 'best MSE and SSIM values' with no numerical margins, error bars, or significance tests. With only 8 test subjects, the claim that this ranking holds 'for all geometries' is not quantified; if the margins are small relative to subject-level variability, the ranking may not be stable. The paper should report per-subject results, confidence intervals or paired statistical tests, and effect sizes for the 2D-versus-2.5D/3D comparison in each geometry.
[Full text (entire submission)] The supplied full text is corrupted and interleaved with an unrelated cs.LG preprint, including federated-learning figures and tables that are not part of this study. As a result, the Methods section, the Results tables, the network architectures, the training hyper-parameters, and the evaluation protocol cannot be verified from the submitted manuscript. This blocks a substantive technical review of the claims. The authors must resubmit a clean, complete, and correctly assembled manuscript before the paper can be evaluated.
[Abstract (2.5D definition)] The 2.5D representation is defined as three orthogonal 64x64 center-plane cuts of each 64x64x64 block arranged as a 64x64x3 image. This discards all off-center voxels and is not a standard multi-planar 2.5D encoding; it is closer to a three-channel patch constructed from one block. The paper should clarify what information is preserved and lost in this representation and why it is expected to encode volumetric context, since the center-plane-only construction may itself limit the volumetric information available to the network.

minor comments (4)

[Abstract] The abstract should include the actual MSE and SSIM values for each geometry and dimensionality condition, or at least refer to a specific results table, so that readers can assess the magnitude of the claimed differences.
[Methods (presumed, unverifiable in corrupted text)] The paper should state explicitly whether the 2D and 3D U-Nets have the same architecture and comparable parameter counts; otherwise the comparison may conflate input dimensionality with model capacity.
[Methods (presumed, unverifiable in corrupted text)] The simulation protocol for sparse sampling should be described precisely, including the number of views used for sparse scans, whether noise was added, and how the full-view reference volumes were generated and registered with the sparse reconstructions.
[Full text] The manuscript contains an unrelated arXiv identifier (arXiv:2508.17954v1) and federated-learning content; this strongly suggests the wrong file or a corrupted build was submitted, and the authors should verify the uploaded PDF.

Circularity Check

0 steps flagged · score 0.0 of 10

Empirical benchmark with external metrics; no circular reduction found.

full rationale

This paper is an empirical benchmark study rather than a derivation. It simulates sparse-view CT data for parallel, fan, and cone beam geometries, trains 2D, 2.5D, and 3D U-Nets, and reports MSE and SSIM on held-out test subjects. The central claim that the 2D axial-slice U-Net achieves the best MSE and SSIM for all geometries is a measured training outcome evaluated against external image-quality metrics; no parameter is fitted to reproduce that conclusion, no uniqueness theorem or prior result by the same authors is invoked to forbid alternatives, and the reported advantage is not forced by construction. The comparison does confound input dimensionality with voxel-block size, resolution, and field of view, since the 2D condition uses full 512x512 slices while the 2.5D and 3D conditions use 64x64x64 blocks. That is a validity threat to the interpretation that dimensionality alone explains the result, but confounding is not circularity: the winning condition is not equivalent to the paper's conclusion by definition, and the result even runs contrary to the authors' stated expectation that volumetric context should help. No equation reduces to its own input, and no load-bearing self-citation chain appears in the abstract or available text. Accordingly, no significant circularity is present, and the appropriate finding is a non-finding with score 0.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

No fitted parameters are reported in the abstract; design constants such as 2048 views, 512x512 slices, 64x64x64 blocks, and the 14/8 subject split are experimental choices, not fitted values. The central claim rests on three domain assumptions: Astra simulations faithfully reproduce clinical sparse-sampling streak artifacts, 2048 views is an adequate full-view ground truth, and the 2D versus 2.5D versus 3D setup isolates dimensionality. The third assumption is the most fragile and is also listed in red_flags and weakest_assumption. The 2.5D scheme is a data layout, not an invented physics entity, so no entities are listed. Architecture and training hyperparameters would belong in this ledger but were not readable.

assumptions (3)

domain assumption Sparse-sampled CT volumes simulated with the Astra toolbox for parallel, fan, and cone beam geometries faithfully reproduce the streak artifacts of real clinical sparse sampling.
All training and test inputs are simulated from retrospectively selected clinical volumes. The abstract does not validate the simulated artifacts against real sparse acquisitions, and artifact physics differ across beam geometries.
domain assumption 2048 views constitute a full-view scan, providing an adequate artifact-free ground truth for training and evaluation.
The abstract fixes 2048 views as the full-view condition without stating the angular sampling scheme or comparing against a denser reference. Residual streaks at 2048 views would corrupt the target labels.
domain assumption Comparing 512x512 axial 2D inputs against 64x64x64-block 2.5D and 3D inputs isolates input dimensionality as the causal variable.
Resolution, field of view, and the network's ability to see whole-image streak artifacts change across conditions. The abstract interprets the outcome as a dimensionality effect without controlling these, which also appears in red_flags and weakest_assumption.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Beam Geometry and Input Dimensionality: Impact on Sparse-Sampling Artifact Correction for Clinical CT with U-Nets." pith.science (2026). https://pith.science/paper/VOH4YK65

@misc{pith2026250817961,
  author       = {Pith},
  title        = {Pith review of: Beam Geometry and Input Dimensionality: Impact on Sparse-Sampling Artifact Correction for Clinical CT with U-Nets},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VOH4YK65}},
  note         = {Machine review of arXiv:2508.17961}
}

read the original abstract

This study aims to investigate the effect of various beam geometries and dimensions of input data on the sparse-sampling streak artifact correction task with U-Nets for clinical CT scans as a means of incorporating the volumetric context into artifact reduction tasks to improve model performance. A total of 22 subjects were retrospectively selected (01.2016-12.2018) from the Technical University of Munich's research hospital, TUM Klinikum rechts der Isar. Sparsely-sampled CT volumes were simulated with the Astra toolbox for parallel, fan, and cone beam geometries. 2048 views were taken as full-view scans. 2D and 3D U-Nets were trained and validated on 14, and tested on 8 subjects, respectively. For the dimensionality study, in addition to the 512x512 2D CT images, the CT scans were further pre-processed to generate a so-called '2.5D', and 3D data: Each CT volume was divided into 64x64x64 voxel blocks. The 3D data refers to individual 64-voxel blocks. An axial, coronal, and sagittal cut through the center of each block resulted in three 64x64 2D patches that were rearranged as a single 64x64x3 image, proposed as 2.5D data. Model performance was assessed with the mean squared error (MSE) and structural similarity index measure (SSIM). For all geometries, the 2D U-Net trained on axial 2D slices results in the best MSE and SSIM values, outperforming the 2.5D and 3D input data dimensions.

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

1 extracted references · 1 canonical work pages

[1]

�� ...

work page Pith review arXiv 2025

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.

[1] [1]

�� ...

work page Pith review arXiv 2025