{"id":"1c6c402f-2d8f-4029-8a92-36611bc8c6da","arxiv_id":"2412.04525","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Using seven contiguous CT slices as input improves super-resolved image quality and defect detection for additively manufactured parts compared with single-slice 2D networks, at near-2D memory cost.","lead":"An Oak Ridge team tested a 2.5D super-resolution trick for X-ray CT scans of 3D-printed metal parts: super-resolve each image slice using its six neighboring slices as context. Across three neural networks, this approach beat single-slice 2D methods on image quality and defect detection while adding under 3% memory use, and 3D methods needed far more memory.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed 2.5D super-resolution gain is measured on a compound task (4x binning, fewer views, and missing beam-hardening correction), so the 2.5D-over-2D PSNR and defect-detection improvements may reflect artifact correction or through-plane interpolation rather than super-resolution per se.","rationale":"The reader's weakest_assumption correctly identifies the compound-degradation confound: the low-resolution input differs from the high-resolution target not only in resolution but also in view count and beam-hardening correction. Section IV-B explicitly acknowledges that the network is also tasked with removing beam-hardening and matching MBIR texture, which confirms that this is not a pure super-resolution experiment. Because the paper's central claim is framed as a super-resolution result, the missing ablation is the single most load-bearing gap. The paper is not internally inconsistent: the methods are clearly described, the memory and parameter tables are concrete, and the 2.5D architecture is a simple extension of established work. The issue is interpretive and empirical: the current evidence cannot distinguish whether 2.5D helps because it better uses neighboring-slice information for actual spatial super-resolution or because it better removes slice-correlated artifacts. A conditional verdict is therefore appropriate, requiring either a resolution-only ablation or a revised claim that explicitly covers joint super-resolution and artifact correction. I do not see grounds to reject: the practical pipeline may still benefit, and the memory comparison is useful. Other weaknesses noted by the reader, such as single test volume and unreleased code, are secondary and do not change the verdict.","tokens_in":8917,"tokens_out":6028,"duration_ms":62634,"concrete_test":"Generate a resolution-only synthetic test set from the same CAD models: reconstruct both low- and high-resolution volumes with the same 2132-view scan, the same MBIR algorithm, and the same beam-hardening correction, differing only in 4x detector binning. Retrain or fine-tune the 2D and 2.5D versions of SRCNN, EDSR, and ESRGAN on this paired data and compare mean PSNR and defect-detection F1. If the 2.5D advantage over 2D persists with a similar margin, the super-resolution claim is supported; if the margin shrinks or reverses, the current headline figures are dominated by artifact correction and view-count mismatch, and the paper's claims must be reframed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Table I defines the synthetic low-resolution input and high-resolution target not only by detector binning (4x) but also by view count (1066 vs 2132) and reconstruction pipeline (uncorrected FDK vs MBIR with beam-hardening correction). The real dataset widens this gap further (147 vs 1066 views). Section IV-B states: 'The job of the networks, apart from super-resolving low-resolution images, is to remove the beam-hardening effect and match MBIR's texture which will enhance defect contrast/detectability.' That sentence is the paper's own admission that the experiment conflates spatial super-resolution with artifact correction, view synthesis, and noise-texture matching. Neighboring slices in XCT carry correlated streak and beam-hardening artifacts, so a 2.5D network is expected to outperform a 2D network on artifact removal even if it provides no true super-resolution benefit. The central claim that 2.5D architectures 'attain a higher mean PSNR than the 2D architectures' and 'consistently attain higher recall, precision, and F1 score' is therefore not yet established for super-resolution specifically. The contribution could stand if reframed as joint 2.5D restoration, but the abstract and conclusions tie it to 'super-resolution.' This is the weakest load-bearing link because every quantitative figure in the paper (Figs. 3, 4, and 5) depends on this compound task, and the stated practical motivation is resolution-limited XCT scanning.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a 2.5D super-resolution architecture for XCT-based inspection of additively manufactured parts. The architecture takes seven contiguous low-resolution slices as input and outputs a super-resolved center slice, and is applied to three standard networks (SRCNN, EDSR, and ESRGAN) by changing only the input channel layout. The authors compare 2D, 2.5D, and 3D variants on a synthetic XCT dataset and present preliminary visual results on a real XCT dataset. The central claims are that 2.5D variants attain higher mean PSNR and higher recall, precision, and F1 scores than 2D variants with a negligible increase in memory, while 3D variants require over 1000% more memory.","tokens_in":9150,"tokens_out":3280,"duration_ms":30729,"significance":"If the central claim holds, the 2.5D architecture is a practically useful bridge between 2D and 3D super-resolution for XCT inspection, because it provides most of the benefit of 3D context at near-2D cost. The paper has several strengths: the comparison is internally consistent (same networks, same patches, only the channel layout changes), the memory and parameter measurements are concrete, and the use of task-specific defect-detection metrics (recall, precision, F1) is more informative than PSNR alone. However, the evidence rests on a single synthetic test volume, a hand-selected window size, and a compound degradation that mixes resolution loss with view-count reduction and missing beam-hardening correction, so the specific claim of 'super-resolution' improvement is not yet established.","major_comments":[{"comment":"The experimental design conflates super-resolution with artifact correction and view synthesis: the low-resolution inputs differ from the high-resolution targets not only by 4x detector binning but also by view count (1066 vs. 2132 for synthetic, 147 vs. 1066 for real) and by reconstruction pipeline (uncorrected FDK vs. MBIR with beam-hardening correction). The manuscript itself states in Section IV-B that 'the job of the networks, apart from super-resolving low-resolution images, is to remove the beam-hardening effect and match MBIR's texture,' which means the reported PSNR and defect-detection gains in Figures 3 and 4 may reflect artifact removal and view interpolation rather than spatial super-resolution. Please either (a) add a controlled synthetic experiment in which the input and target differ only by binning (same views, same reconstruction pipeline, matched or no beam-hardening correction) and re-run the comparisons, or (b) explicitly reframe the contribution as joint 2.5D restoration and temper the super-resolution-specific claims in the abstract and conclusions.","section":"IV (Datasets, Section IV-B)"},{"comment":"The choice of seven input slices is reported as 'selected because it gave the best results' without stating the validation protocol, the range of window sizes tested, or the performance for other window sizes. Since all quantitative comparisons are reported on a single synthetic test volume, selecting the window size on that same test data introduces a selection bias that is not quantified. Please report the window-size search on a held-out validation set and provide the PSNR and defect-detection metrics for at least one other window size.","section":"IV (2.5D Super-Resolution, window size selection)"},{"comment":"The real-data evidence is visual and limited to ESRGAN, with no quantitative PSNR or defect-detection metrics, yet the conclusion states that 'preliminary results on real data show that 2.5D outperforms 2D super-resolution and performs similarly to 3D super-resolution.' That statement is stronger than the evidence supports. Please either add quantitative real-data evaluation or soften the conclusion to something like 'qualitatively, on a single test volume, 2.5D appears to improve defect visibility compared to 2D and to approach 3D.'","section":"IV-C (Preliminary Results on Real Dataset)"},{"comment":"The synthetic comparison is based on one test volume, and the standard deviations in Figure 3 are computed over slices of that volume rather than over independent scans, reconstructions, or network training runs. This makes it difficult to assess whether the reported 2.5D-over-2D improvements are statistically meaningful. Please evaluate on multiple test volumes (for example, different CAD geometries, noise realizations, or scan parameters) and report per-volume or per-seed variability.","section":"IV-B (Results on Synthetic Dataset)"}],"minor_comments":[{"comment":"In the introduction, 'a essential tool' should be 'an essential tool.'","section":"Section I"},{"comment":"The section numbering is confusing: Section III is titled 'Experimental Results' but contains dataset and architecture subsections (III-A, III-B), while Section IV is titled '2.5D Super-Resolution' and contains the actual implementation and results. Please renumber so that methods appear before results.","section":"Sections III and IV"},{"comment":"In the synthetic dataset description, 'python's spekpy package' should be capitalized as 'Python's SpekPy package' for consistency with the package name.","section":"Section IV-A"},{"comment":"The caption states 'as denoted by the arrows,' but the figure as provided does not show arrows in the text; please ensure the arrows are visible in the final figure or describe the indicated regions in words.","section":"Figure 5 caption"}],"recommendation":"major_revision","confidential_remarks":"The paper is a straightforward empirical engineering study with modest novelty; the 2.5D architecture alteration itself is simple but the memory/parameter comparison is useful. The main risk is the compound-degradation issue described in Major Comment 1: if the authors cannot add a binning-only controlled experiment, the paper should be reframed as 2.5D restoration rather than super-resolution. The self-citations to prior 2.5D work are contextual and not problematic."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read it. The paper does something simple and useful: it takes standard 2D SR networks (SRCNN, EDSR, ESRGAN), widens the first layer to accept seven contiguous slices, and shows on synthetic XCT volumes that this 2.5D variant beats 2D on PSNR and on defect detection recall/precision/F1, with under 3% extra memory, while 3D costs >1000% memory. That empirical comparison is new for XCT of AM parts, even though the 2.5D trick itself is not (they cite prior 2.5D reconstruction/denoising/segmentation work). The experiments look internally consistent: same patch sizes, same training, only input channels differ. The memory and parameter tables are useful. Credit where due: the paper is honest about what the network is being asked to do.\n\nThe soft spot is the one the authors themselves flag in Section IV-B: the low-resolution inputs differ from the high-resolution targets in three ways — 4x binning, half (or fewer) the views, and no beam-hardening correction on FDK vs MBIR with correction. So the network is simultaneously super-resolving, denoising, view synthesizing, and removing beam hardening. The 2.5D advantage could come from using neighboring slices to clean up correlated streak artifacts rather than from true super-resolution. The paper calls this a super-resolution study, but really it's a joint restoration study. The fix is straightforward: repeat the comparison with matched view counts and reconstruction pipelines, or add an ablation where only binning changes. Without that, the specific claim that 2.5D helps for super-resolution per se is not isolated.\n\nOther soft spots are minor: the 7-slice window was chosen for best results with no described validation protocol; metrics come from a single synthetic test volume with no run-to-run variability; real-data results are visual and ESRGAN-only; no code or data released. These are addressable.\n\nFor peer review: yes, I would send it. The work is relevant to industrial XCT, the comparisons are fair as far as they go, and the main confound is explicitly acknowledged. A careful reviewer can push for the ablation and release of code. The paper deserves a revision rather than a desk reject.","headline":"A practical 2.5D SR study for XCT of AM parts with solid baselines, but the claimed SR gain is entangled with artifact correction.","tokens_in":9833,"tokens_out":2044,"would_cite":false,"duration_ms":19936,"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":"A 2.5D super-resolution architecture—seven neighboring CT slices in, one super-resolved center slice out—beats 2D methods on XCT defect detection while adding less than 3 percent memory, making most of 3D's benefit available at near-2D…","keywords":["X-ray computed tomography","super-resolution","2.5D deep learning","additive manufacturing","non-destructive evaluation","defect detection","deep learning","MBIR"],"falsifier":"Re-run the same 2D, 2.5D, and 3D training with input and target volumes reconstructed by the same algorithm so that only detector binning and view count change one at a time; if the 2.5D advantage over 2D on recall, precision, and F1 shrinks or disappears, the reported gain is driven by artifact and texture matching rather than by inter-slice resolution recovery.","tokens_in":8635,"feed_emoji":"🩻","tokens_out":6686,"duration_ms":60089,"temperature":0.7,"pith_summary":"This paper tries to establish that a simple architecture change—feeding seven contiguous low-resolution CT slices into a super-resolution network and asking it to output the super-resolved center slice—gives most of the benefit of full 3D processing at nearly 2D cost. On synthetic XCT volumes of additively manufactured aluminum parts, the 2.5D versions of SRCNN, EDSR, and ESRGAN all achieve higher mean PSNR than their 2D counterparts while increasing memory by less than 3 percent, and they consistently improve defect-detection recall, precision, and F1 score. Full 3D versions of the same networks score highest on image quality and defect detection but require over 1000 percent more memory, which the paper argues makes them impractical for large-scale inspection. The result matters because high-resolution XCT of metal parts is slow and expensive, and faster low-resolution scans lose the small defects that determine part quality; if the 2.5D advantage holds, it offers a practical route to high-throughput, accurate non-destructive evaluation.","feed_headline":"2.5D super-resolution lifts CT defect detection at 2D cost","feed_subtitle":"Feeding seven CT slices into a super-resolution net beats 2D on defect detection while adding under 3% memory.","key_machinery":"The load-bearing mechanism is the 2.5D input modification: seven contiguous low-resolution slices are concatenated as input channels to an otherwise unmodified 2D super-resolution network, and the loss is computed only on the output center slice. Because the extra six input channels are absorbed by the first convolutional layer, the parameter increase is a constant $6mnk$ (kernel height $m$, kernel width $n$, and $k$ output feature maps), independent of patch size, which is why memory grows by less than 3 percent. The sliding-window scheme then assembles a full super-resolved volume while preserving the computational footprint of 2D inference, giving the network access to through-plane spatial context without 3D convolutions.","core_discovery":"The central discovery is that inter-slice context, not full volumetric processing, is what makes super-resolution useful for XCT defect detection. The paper proposes a generic 2.5D architecture: take any 2D super-resolution network and change its input from a single slice to seven contiguous low-resolution slices, with the network trained to output the super-resolved center slice; a sliding window of seven slices then produces the whole volume. In synthetic experiments on aluminum AM parts, 2.5D architectures attain higher mean PSNR than 2D architectures with a negligible increase in required memory (2.63 percent for SRCNN, 0.10 percent for EDSR, 0.005 percent for ESRGAN), while 3D architectures increase memory by 1,122 to 12,416 percent. On defect detection, 2.5D consistently attains higher recall, precision, and F1 than 2D, with the largest gains in recall, and the paper emphasizes that 2D methods look acceptable on PSNR yet fail badly on task-specific defect metrics. Preliminary real-data results with ESRGAN show 2.5D resolving small defects slightly better than 2D and producing fewer Z-direction artifacts, visually on par with 3D for small defects.","pith_inferences":["Because the synthetic and real training pairs differ in view count, detector binning, and beam-hardening correction simultaneously, part of the measured gain may come from learning to remove beam-hardening artifacts and match MBIR's noise texture rather than from recovering true spatial resolution; a controlled ablation varying one degradation at a time would separate these effects.","The 7-slice window was selected empirically, which suggests the optimal window size is a tunable parameter that could be re-tuned for anisotropic voxels or different defect sizes, potentially changing the memory-quality trade-off.","The paper tests only ESRGAN on the real dataset; a natural extension is to check whether the 2.5D advantage persists for SRCNN and EDSR on real scans, since the paper reports those struggle on real data even in 2D.","The architecture is not specific to XCT or to these three backbones, so the same 2.5D modification could be applied to transformer- or diffusion-based super-resolution models, which the paper lists as future work."],"forward_implications":["The 2.5D input trick can be dropped into any existing 2D super-resolution network; the paper demonstrates it on SRCNN, EDSR, and ESRGAN with the same standard architectures.","Defect detection on XCT volumes improves over 2D super-resolution on recall, precision, and F1, with recall improving most, meaning small defects that 2D methods miss become detectable.","Full 3D super-resolution remains the quality leader, but its memory cost (over 1000 percent more than 2D) makes it impractical for large industrial volumes, so 2.5D is the viable middle ground.","On real steel XCT data, preliminary ESRGAN results indicate 2.5D reduces Z-direction artifacts compared to 2D and visually matches 3D on small-defect recovery, despite 3D suffering from stitching grid artifacts under memory limits.","The approach targets high-throughput non-destructive evaluation of additively manufactured parts, where scanning time and cost push toward low-resolution acquisition."],"supporting_citations":[{"why":"Supplies the SRCNN baseline network that is adapted into a 2.5D architecture and tested against its 2D and 3D versions.","marker":"[6]"},{"why":"Supplies the EDSR residual-network baseline whose 2.5D variant is compared on PSNR and defect detection.","marker":"[10]"},{"why":"Supplies the ESRGAN generator and discriminator baseline used for both synthetic and real-data experiments.","marker":"[12]"},{"why":"Prior demonstration that 2.5D deep learning for CT reconstruction balances image fidelity with computational cost, motivating the architecture.","marker":"[14]"},{"why":"Empirical study of 2.5D volumetric segmentation supporting the claim that 2.5D approaches can approach 3D quality at lower compute.","marker":"[21]"},{"why":"FDK reconstruction algorithm used to create the low-resolution input volumes from sparse or binned projection data.","marker":"[24]"},{"why":"Model-based iterative reconstruction used to generate the high-resolution target volumes for training and testing.","marker":"[25]"},{"why":"Beam-hardening correction applied to the high-resolution targets, making artifact removal part of the super-resolution task.","marker":"[27]"}],"fun_headline_variants":["2.5D super-res uses slice context to boost CT defect detection","Seven-slice window sharpens XCT defect detection at 2D memory cost","Inter-slice context, not full 3D, sharpens CT defect detection","2.5D super-resolution: 2D speed, 3D-like CT defect clarity"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison relies on treating the high-resolution model-based iterative reconstructions with beam-hardening correction as the true target; if those targets carry their own artifacts, the measured PSNR and defect-detection gains may reflect learning to imitate a particular reconstruction algorithm rather than recovering genuine spatial detail.","fun_headline_variants_meta":{"raw":{"variants":["2.5D super-res uses slice context to boost CT defect detection","Seven-slice window sharpens XCT defect detection at 2D memory cost","Inter-slice context, not full 3D, sharpens CT defect detection","2.5D super-resolution: 2D speed, 3D-like CT defect clarity"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000629,"raw_usage":{"total_tokens":2964,"prompt_tokens":1062,"completion_tokens":1902,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":678,"completion_tokens_details":{"reasoning_tokens":1815}},"tokens_in":678,"tokens_out":1902,"duration_ms":14261,"temperature":1.0,"reasoning_tokens":1815,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T21:33:33.895309+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the same 2D, 2.5D, and 3D training with input and target volumes reconstructed by the same algorithm so that only detector binning and view count change one at a time; if the 2.5D advantage over 2D on recall, precision, and F1 shrinks or disappears, the reported gain is driven by artifact and texture matching rather than by inter-slice resolution recovery.","supporting_citations":[{"cited_title":"Enhanced deep residual networks for single image super-resolution,","cited_arxiv_id":null,"evidence_quote":"Supplies the EDSR residual-network baseline whose 2.5D variant is compared on PSNR and defect detection."},{"cited_title":"Esrgan: Enhanced super-resolution generative adversar- ial networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the ESRGAN generator and discriminator baseline used for both synthetic and real-data experiments."},{"cited_title":"2.5D deep learning for CT image reconstruction using a multi-gpu implementation,","cited_arxiv_id":null,"evidence_quote":"Prior demonstration that 2.5D deep learning for CT reconstruction balances image fidelity with computational cost, motivating the architecture."},{"cited_title":"Bridging 2D and 3D seg- mentation networks for computation-efficient volumetric medical image segmentation: An empirical study of 2.5D solutions,","cited_arxiv_id":null,"evidence_quote":"Empirical study of 2.5D volumetric segmentation supporting the claim that 2.5D approaches can approach 3D quality at lower compute."},{"cited_title":"Practical cone-beam algorithm,","cited_arxiv_id":null,"evidence_quote":"FDK reconstruction algorithm used to create the low-resolution input volumes from sparse or binned projection data."},{"cited_title":"Fast model-based X-ray CT reconstruction using spatially nonhomogeneous ICD optimization,","cited_arxiv_id":null,"evidence_quote":"Model-based iterative reconstruction used to generate the high-resolution target volumes for training and testing."},{"cited_title":"Neural Network-based Single-material Beam Hardening Correction for X-ray CT in Addi- tive Manufacturing ,","cited_arxiv_id":null,"evidence_quote":"Beam-hardening correction applied to the high-resolution targets, making artifact removal part of the super-resolution task."}],"review_version":1}