{"id":"1fb65710-f748-4ee2-92ca-c8f3e70db4af","arxiv_id":"2508.17961","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"On 22 clinical CT volumes with simulated sparse sampling, the 2D axial-slice U-Net scored best on MSE and SSIM for parallel, fan, and cone beam geometries, beating tri-planar 2.5D patches and 64x64x64 3D blocks.","lead":"A Munich hospital study of 22 clinical CT scans found that a plain 2D U-Net removes sparse-sampling streak artifacts better than 2.5D or 3D input variants, across all three tested beam geometries. The result is a practical caution for CT artifact correction: volumetric context does not automatically help, and simpler 2D pipelines may remain the default.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 2D-vs-2.5D/3D comparison confounds input dimensionality with spatial resolution and field of view: 2D uses full 512x512 slices while volumetric variants use 64x64x64 blocks, so the reported 2D advantage does not establish that dimensionality is the operative factor.","rationale":"The reader's weakest_assumption identifies exactly the confound I consider load-bearing: the comparison does not isolate dimensionality because the 2D condition uses full-resolution 512x512 axial slices while the 2.5D and 3D conditions use small 64x64x64 blocks. This is not a matter of external consensus or author intent; it is an internal design property described in the abstract's preprocessing paragraph. The paper's stated purpose is to evaluate the effect of input dimensionality on artifact correction, yet the experiment varies resolution, field of view, and available anatomical context together with dimensionality. A 64-voxel block is too small to contain a whole streak artifact, so the volumetric models are handicapped in a way that has nothing to do with whether context is 2D or 3D. Conversely, the 2D model's full-slice input can see the global streak pattern. The headline claim therefore cannot be accepted on the presented evidence. I credit the study for a coherent simulation pipeline and a stated hypothesis that the result contradicts, so there is no circularity concern; the issue is construct validity. The full text supplied to me is corrupted and interleaved with an unrelated preprint, so I cannot rule out the possibility that the full paper includes a matched-resolution control or additional analyses. That uncertainty is precisely why the reader's UNVERDICTED verdict should stand: my concern reinforces the need for a full review and a matched comparison rather than moving the verdict to accept or reject. If the full text already contains such a control, this objection would be resolved; if not, the abstract's conclusion should be weakened to a pipeline-level comparison, not a dimensionality finding.","tokens_in":19965,"tokens_out":4601,"duration_ms":46782,"concrete_test":"Extract the 64x64 center patches from the full axial slices at exactly the locations used to form the 2.5D and 3D blocks, and retrain the same 2D U-Net architecture on those patches with an identical training schedule. If the patched 2D model no longer matches or beats the 2.5D and 3D models on test MSE and SSIM, the reported 2D advantage is attributable to full-slice resolution and field of view rather than to input dimensionality.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim, from the abstract, is that 'for all geometries, the 2D U-Net trained on axial 2D slices results in the best MSE and SSIM values,' and the study is framed as isolating 'dimensions of input data' as a way of incorporating volumetric context. For that conclusion to hold, the comparison must vary only input dimensionality while holding resolution, field of view, and available information constant. The preprocessing described in the abstract does not do this. The 2D condition trains on full 512x512 axial slices, while the 2.5D and 3D conditions are built from 64x64x64 voxel blocks; 2.5D keeps only three 64x64 center-plane cuts rearranged as a 64x64x3 image, and 3D uses individual 64-voxel blocks. Thus the volumetric conditions simultaneously reduce spatial resolution, shrink the field of view, and lose the global image context, while the winning 2D condition sees the full slice. Sparse-sampling streak artifacts in CT are global: they propagate across the full field of view, so a 64x64 block cannot represent them, whereas a 512x512 slice can. The reported 2D advantage is therefore fully explainable by information loss in the block representation, not by dimensionality. Additionally, with only 8 test subjects and no error bars or significance tests reported in the abstract, the consistency of the advantage is not quantified. The supplied full text is corrupted and interleaved with an unrelated cs.LG preprint, so I could not verify whether the full paper contains additional matched controls; based on the available evidence, the central claim is underdetermined.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":20083,"tokens_out":3896,"duration_ms":40406,"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":[{"comment":"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.","section":"Abstract (pre-processing description)"},{"comment":"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.","section":"Abstract (results)"},{"comment":"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.","section":"Full text (entire submission)"},{"comment":"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.","section":"Abstract (2.5D definition)"}],"minor_comments":[{"comment":"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.","section":"Abstract"},{"comment":"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.","section":"Methods (presumed, unverifiable in corrupted text)"},{"comment":"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.","section":"Methods (presumed, unverifiable in corrupted text)"},{"comment":"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.","section":"Full text"}],"recommendation":"major_revision","confidential_remarks":"The submitted full text is corrupted and interleaved with an unrelated preprint, so the editor should confirm that the correct PDF was uploaded. Even after cleaning up the text, the central empirical claim needs a matched-resolution comparison or a substantially reframed conclusion; the current study does not support the stated role of input dimensionality. I recommend major revision rather than rejection because the confound can in principle be addressed with additional experiments or a narrower claim, and the practical question is worth pursuing."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","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.\n\nWhat'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.\n\nThe 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.\n\nI 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.\n\nBottom 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.\"\n\nRecommendation: 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.","headline":"The 2D advantage is plausible but the comparison confounds dimensionality with resolution and field of view; the geometry sweep is still a useful benchmark.","tokens_in":20862,"tokens_out":2430,"would_cite":false,"duration_ms":24228,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"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.","keywords":["sparse-sampling CT","streak artifact correction","U-Net","input dimensionality","2.5D input","3D U-Net","beam geometry","clinical CT"],"falsifier":"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.","tokens_in":19565,"feed_emoji":"🩻","tokens_out":7984,"duration_ms":73653,"temperature":0.7,"pith_summary":"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.","feed_headline":"2D beats 2.5D and 3D for CT streak-artifact correction","feed_subtitle":"Across parallel, fan, and cone beam scans, the full 512x512 axial-slice U-Net wins on MSE and SSIM.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[],"fun_headline_variants":["2D wins over 2.5D and 3D for CT streak removal","Counterintuitive: 2D U-Net beats volumetric inputs for CT artifacts","For CT sparse-sampling streaks, 2D beats 3D context","Volumetric context hurts: 2D U-Net top for CT artifact correction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["2D wins over 2.5D and 3D for CT streak removal","Counterintuitive: 2D U-Net beats volumetric inputs for CT artifacts","For CT sparse-sampling streaks, 2D beats 3D context","Volumetric context hurts: 2D U-Net top for CT artifact correction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000289,"raw_usage":{"total_tokens":1761,"prompt_tokens":1081,"completion_tokens":680,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":697,"completion_tokens_details":{"reasoning_tokens":592}},"tokens_in":697,"tokens_out":680,"duration_ms":6428,"temperature":1.0,"reasoning_tokens":592,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:58:59.187053+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[],"review_version":2}