{"id":"48d61f14-e82b-46b2-8d4f-c1eec5ccf451","arxiv_id":"2505.03412","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"CXR-AD is a new X-ray benchmark for internal component defects, with 653 normal and 561 defect images across five component categories; three standard anomaly detection methods all show substantial performance drops compared to MVTec AD.","lead":"This paper presents CXR-AD, a new X-ray image dataset of electronic components with pixel-level labels of internal defects. It benchmarks three standard anomaly-detection algorithms and finds they perform far worse on these X-ray images than on the usual MVTec AD dataset, which matters for automated quality control in manufacturing.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The manuscript provides no URL or repository for CXR-AD, so the abstract's claim that it is the 'first publicly accessible' component X-ray anomaly detection dataset is unsupported and the benchmark cannot be independently reproduced.","rationale":"I read the paper as a dataset and benchmark contribution. For that central claim to hold, the data must be obtainable; the manuscript gives no path. The paper does provide positive evidence: real X-ray acquisition details (YXLON Cheetah, 16-bit, controlled environment), quantitative image statistics (average local contrast 0.1091, 95.47% of pixels in the 50–150 range), and a benchmark on three representative methods (PatchCore, ONENIP, AdaCLIP). These are useful and suggest the dataset exists, but they do not establish public accessibility. The reader's weakest assumption about mask fidelity is also valid, but a missing dataset makes the mask question moot for external validation. I therefore recommend keeping the CONDITIONAL verdict, with the release mechanism as the explicit condition. If the authors provide a working data link, the main objection is resolved; then the mask ground truth should be tested by independent re-annotation of a random subset.","tokens_in":9859,"tokens_out":5566,"duration_ms":56439,"concrete_test":"Search the manuscript and standard scholarly data repositories (Zenodo, GitHub, IEEE DataPort, the SYSU author pages) for any CXR-AD download link or DOI. If no public dataset artifact can be located, the 'publicly accessible' claim in the Abstract is not supported; the benchmark results and mask-quality statements cannot be independently verified until release. This one check is sufficient to settle the concern.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is in the Abstract: 'we construct the first publicly accessible component X-ray anomaly detection (CXR-AD) dataset.' No download link, hosting service, or data-availability statement appears anywhere in Sections I–V, the footnotes, or the references. A dataset paper without a release artifact cannot support the word 'publicly accessible,' and without the data the quantitative benchmark (Table II) and the 'precise pixel-level mask annotations' cannot be checked by other researchers. The reader's concern about mask annotation quality is real: masks come from one described LabelMe workflow, with vertex adjustments repeated 5–8 times per sample and checked only by an automated 'geometrical consistency' module (Section III), which does not validate the semantic correctness of defect boundaries. However, this problem is secondary because it assumes the data can be obtained; currently it cannot. Additionally, the 'first' claim is left unreconciled with earlier public X-ray inspection datasets such as GDXray, which contains X-ray images of castings with defects; even if the authors intend to restrict the claim to semiconductor components, the manuscript does not state that delimitation when making the 'first' claim. Both issues are correctable, but the missing release mechanism is the most immediate load-bearing gap.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper introduces CXR-AD, an industrial X-ray image dataset for anomaly detection in semiconductor components. It reports 653 normal images (559 train + 94 test normal) and 561 defective test images across five classes, with pixel-level polygonal masks produced with LabelMe. The authors characterize the data using grayscale histograms, local contrast, and defect-scale distributions, and benchmark PatchCore, ONENIP, and AdaCLIP in image- and pixel-level AUROC. They report an average image-level performance degradation of 29.78% relative to MVTec AD, and conclude that existing algorithms are limited for internal defect detection.","tokens_in":10128,"tokens_out":6847,"duration_ms":67625,"significance":"The dataset addresses a genuine gap: most industrial anomaly-detection benchmarks are visible-light surface-defect datasets, whereas X-ray inspection of internal component defects is an important real-world task. The benchmark results are informative, and the acquisition setting (microfocus X-ray system with 16-bit dynamic range) is realistic. The central performance-degradation claim is supported by Table II and is not circular relative to the dataset's descriptive statistics. Provided the dataset is actually released, CXR-AD could become a useful testbed. At present, however, the absence of a public release mechanism makes the 'publicly accessible' claim unverifiable, and several dataset-characterization statistics need correction.","major_comments":[{"comment":"The central claim that CXR-AD is the 'first publicly accessible component X-ray anomaly detection dataset' is unsupported because the manuscript contains no repository URL, download link, hosting service, or data-availability statement anywhere in Sections I–V, the footnotes, or the references. Without a release mechanism, the benchmark in Table II cannot be reproduced by other researchers and the claim of public accessibility cannot be verified. The authors should either provide the artifact location or qualify the claim in the abstract and introduction.","section":"Abstract; Section III"},{"comment":"The related-work review does not discuss existing public X-ray defect datasets such as GDXray, which contains X-ray images of castings with defects, and the 'first' claim is not delimited at the point where it is made. If the intended scope is specifically semiconductor packaging components, that restriction should be stated explicitly and the differences from prior X-ray inspection benchmarks should be explained; otherwise the novelty claim is overstated.","section":"Section II.A"},{"comment":"The dataset-characterization statistics are not internally consistent and are over-generalized. The local-contrast analysis is computed on 62 samples of the UB class only, yet the text concludes that 'the defective regions in semiconductor X-ray images exhibit extremely low contrast characteristics'; the defect-scale statement that small defects account for 87% while medium and large together account for 12% does not sum to 100; and the bin definitions (1%, 10%, 50% of the maximum defect area) are not specified. Please report per-class statistics or clearly state that the analysis is exemplary, correct the percentages, and define the adaptive binning procedure.","section":"Section III; Figs. 6–8"},{"comment":"The pixel-level mask annotations are created by a single LabelMe-based labeling workflow, with vertex adjustments repeated 5–8 times per sample, and are checked only by an automated geometrical-consistency module. This does not verify the semantic correctness of the defect boundaries, which matters because pixel-level AUROC is one of the paper's main benchmark outputs. The paper should report annotation reliability, for example through a second-annotator study or expert review of a sample of masks, or should explicitly acknowledge this limitation in the conclusion.","section":"Section III; Table II"}],"minor_comments":[{"comment":"The caption of Fig. 7 repeats 'Distribution of local contrast statistics' from Fig. 6, even though the text describes Fig. 7 as showing the defect-scale distribution; the caption should be corrected.","section":"Fig. 7"},{"comment":"The dataset is called 'the Chip X-ray anomaly detection dataset' at the start of Section III, while the title and abstract use 'Component X-ray'; the terminology should be aligned throughout.","section":"Section III"},{"comment":"The claim of a '29.78% average performance degradation' should specify that this is the mean image-level AUROC degradation across the three methods; the corresponding pixel-level average degradation is different and should be stated separately if used.","section":"Abstract; Section IV"},{"comment":"The sentence 'AdaClip underperforms compared to the other two models, suggesting that it may rely heavily on precise textual prompts for specific detection tasks' is speculative; either provide supporting evidence or rephrase as a hypothesis for future work.","section":"Section IV.C"},{"comment":"Reference [21] is cited as 'Language models are few-shot learners' but is invoked as 'ChatGPT'; the citation should be corrected to the appropriate GPT/ChatGPT reference, and reference [13] should be checked because the text describes f-AnoGAN rather than the original AnoGAN.","section":"References"},{"comment":"The paragraph describing mask conversion contains a duplicated statement about batch-converting JSON annotations to binary masks; the repetition should be removed.","section":"Section III"}],"recommendation":"major_revision","confidential_remarks":"The deciding issue is the missing data-release artifact: without a repository link the words 'publicly accessible' cannot be supported, and reviewers cannot verify the masks or reproduce Table II. If the authors supply a working dataset link and resolve the prior-art question regarding GDXray, the paper appears salvageable and the benchmark claim is defensible. I would also ask the editor to encourage the authors to add annotation-reliability evidence, since pixel-level evaluation is central to the paper's contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe short version: CXR-AD is a plausible new dataset artifact and the benchmark gap is real, but the manuscript cannot support its headline \"first publicly accessible\" claim because no repository, URL, or data availability statement appears anywhere in the paper. That is the load-bearing issue.\n\nWhat is actually new: real X-ray images of five semiconductor packaging types (653 normal, 561 defect) with pixel-level masks, captured with a YXLON system from industrial production. The benchmark runs three representative anomaly detection methods (PatchCore, ONENIP, AdaCLIP) and shows a sharp drop relative to MVTec AD—image AUROC around 56–85 vs 99, pixel around 71–94 vs 98. That result supports the claim that internal X-ray inspection is harder than the surface-defect benchmarks. The dataset, if released, would fill a genuine gap.\n\nWhere the soft spots are:\n\n- No release mechanism. The abstract says \"publicly accessible\" twice, but Sections I–V contain no link, hosting service, repository, or availability statement. Without the data, Table II cannot be reproduced and the \"precise masks\" cannot be checked. This is not a minor omission; it is what makes the paper's central claim true or false.\n\n- The \"first\" claim is undercut by prior public X-ray inspection datasets such as GDXray (castings with defects). The authors may mean \"first X-ray dataset for anomaly detection benchmarking of semiconductor components,\" but they never say that. They should cite and differentiate.\n\n- Mask quality rests on a single manual LabelMe workflow, with 5–8 vertex adjustments per defect, checked only by an automated geometric consistency module. That check does not validate semantic correctness of defect boundaries. For pixel-AUROC evaluation, systematic mask errors would bias comparisons. This is a real concern, though not fatal if the masks are re-inspected or a second labeler is added.\n\n- Minor: the local contrast analysis uses only 62 UB samples but is presented as a dataset-wide characterization; defect scale percentages (87% + 12%) do not sum to 100; Fig. 7's caption duplicates the contrast title. These are easy fixes.\n\nOn the math and methods: the benchmark is standard and the performance numbers are internally consistent as far as I can tell. No fitted parameters or circular reasoning drive the conclusion.\n\nThis paper deserves a serious referee—the artifact and the problem are real. But the referee should require a data-availability statement and reconciliation with prior X-ray datasets before publication. I would take it to reading group to discuss the release issue, though I would not cite it yet until the data is actually accessible.","headline":"A useful new dataset benchmark, but the 'publicly accessible' claim is unverifiable without a release mechanism and under-reconciled with prior X-ray data.","tokens_in":10634,"tokens_out":2361,"would_cite":false,"duration_ms":23103,"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":"CXR-AD, the first public component X-ray anomaly detection dataset, causes current anomaly detectors to lose about 30 percent average AUROC compared with their MVTec AD results.","keywords":["X-ray imaging","anomaly detection","industrial defect detection","semiconductor components","dataset benchmark","pixel-level annotation","low-contrast defect","multi-scale anomaly"],"falsifier":"Independently re-annotate a random subset of CXR-AD defect images with several trained annotators and compare the new masks against the published ones with IoU and boundary distance; also have an expert verify the geometric-consistency check on the original images. If inter-annotator agreement is low or the automated check misses misaligned contours, the pixel-level benchmark numbers should not be taken at face value.","tokens_in":9705,"feed_emoji":"🔍","tokens_out":7087,"duration_ms":62966,"temperature":0.7,"pith_summary":"The paper introduces CXR-AD, the first publicly available X-ray dataset for anomaly detection in semiconductor components. It is built from real industrial inspection images: 653 normal and 561 defective radiographs across five chip packaging types, with pixel-level masks on every defect. The authors argue that existing datasets cover only surface defects in visible light, so CXR-AD fills a missing regime where defects are internal, low-contrast, noisy, and highly varied in size. To show the regime is genuinely harder, they benchmark a feature-based method, a reconstruction-based method, and a zero-shot multimodal method, reporting an average 29.78% AUROC degradation relative to MVTec AD. If the dataset holds up, it gives the field a realistic testbed for internal defect inspection and a concrete target: algorithms must improve on low-contrast, multi-scale X-ray anomalies.","feed_headline":"First public X-ray defect dataset cuts anomaly scores 30%","feed_subtitle":"Real internal chip X-rays with pixel masks expose how current detectors fail on low-contrast, tiny defects","key_machinery":"The load-bearing object is the dataset itself, built from a YXLON Cheetah microfocus X-ray system and annotated with LabelMe-drawn polygon contours converted to binary masks. The argument is carried by three quantitative characterizations: a grayscale histogram showing that 95.47% of pixels lie between gray levels 50 and 150; a local-contrast statistic $\\text{LC} = (\\mu_{\\text{obj}} - \\mu_{\\text{bg}})/\\mu_{\\text{bg}}$ averaging 0.1091 across the UB samples; and a defect-scale analysis in which small defects account for 87% of all defects. These measurements operationalize the claimed challenges of low contrast, high noise, and multi-scale anomalies, and they explain why the benchmarked methods lose performance on CXR-AD; in particular, the paper attributes ONENIP's poor pixel-level showing to boundary blur and leakage of small defects.","core_discovery":"The central claim is that CXR-AD is the first public component-level X-ray anomaly detection benchmark and that it is measurably harder than the de facto surface-defect benchmark MVTec AD. The dataset contains five semiconductor packaging classes (CDIP, CFP16, CFP20, UB, DC); 559 normal images form the training set, and the test set has 94 normal plus 561 defective images with pixel-level polygon masks. Quantitative analysis of the UB class shows 95.47% of pixel values in the 50–150 gray range, an average local contrast of 0.1091 between defect and background, and 87% of defects classed as small; the paper presents these as the mechanisms behind the three named challenges: structure–defect coupling, low contrast with noise, and multi-scale defect morphology. Benchmark results on image/pixel AUROC are PatchCore 56.58/74.78, ONENIP 85.34/93.66, and AdaCLIP 55.32/70.80, versus their MVTec averages of 99.00/98.00, 97.90/97.90, and 89.71/89.90, which the paper summarizes as a 29.78% average degradation.","pith_inferences":["The 29.78% gap is computed against each method's own MVTec numbers; a fairer difficulty comparison would also control for test-set composition and image resolution, since X-ray images have different statistics than natural-light surface images.","A concrete next experiment the paper does not run: apply contrast enhancement or multi-scale feature fusion to the three baselines and measure how much of the MVTec gap closes; that would isolate whether the degradation is inherent to X-ray statistics or to current architectures.","The mask-quality assumption could be tested by re-annotating a random subset of CXR-AD defects with multiple annotators and reporting inter-annotator IoU; without that, pixel-AUROC comparisons across future methods may be comparing noise.","The 'first public X-ray component anomaly detection dataset' claim is scope-dependent; if adjacent X-ray inspection datasets exist, the distinctive contribution is the combination of semiconductor components, real production defects, and pixel-level masks."],"forward_implications":["CXR-AD provides a public benchmark on which anomaly detection methods can be compared for internal X-ray inspection, alongside surface-defect benchmarks.","Because all three method families degrade, progress on CXR-AD requires handling low contrast and small defect scales, not merely scaling up existing feature backbones.","The biggest gap is localization: AdaCLIP's pixel AUROC falls to 70.80% and ONENIP's outputs blur defect boundaries, so pixel-level accuracy on X-ray images is the open problem the dataset highlights.","The 559-image normal-only training split supports unsupervised and few-shot anomaly detection research without defective training samples."],"supporting_citations":[{"why":"Supplies the MVTec AD benchmark whose average image/pixel AUROC scores are the reference for the reported 29.78% degradation.","marker":"[2]"},{"why":"Provides the VisA surface-defect dataset used to train the AdaCLIP zero-shot baseline and represents the existing surface-oriented dataset regime.","marker":"[3]"},{"why":"Shows the metal-surface defect focus of earlier industrial anomaly detection datasets, supporting the gap claim.","marker":"[4]"},{"why":"Defines the feature-embedding baseline PatchCore; its CXR-AD image AUROC drops to 56.58 from 99.00 on MVTec AD.","marker":"[19]"},{"why":"Defines the reconstruction-based baseline ONENIP, the best performer on CXR-AD (85.34/93.66) yet still below its MVTec result.","marker":"[27]"},{"why":"Defines the zero-shot multimodal baseline AdaCLIP; its pixel AUROC falls to 70.80 on CXR-AD, showing the localization challenge.","marker":"[26]"}],"fun_headline_variants":["First public X-ray defect dataset is 30% harder for AI","X-ray anomaly benchmark: internal defects cut AI scores 30%","CXR-AD: first X-ray internal defect set, 30% tougher than MVTec","New X-ray dataset exposes 30% drop in anomaly detection accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pixel-level masks, drawn by a single manual LabelMe workflow and checked only by an automated geometric-consistency module, are assumed accurate enough to serve as ground truth; if they contain systematic errors, the reported pixel-AUROC results and the dataset's quality claims collapse.","fun_headline_variants_meta":{"raw":{"variants":["First public X-ray defect dataset is 30% harder for AI","X-ray anomaly benchmark: internal defects cut AI scores 30%","CXR-AD: first X-ray internal defect set, 30% tougher than MVTec","New X-ray dataset exposes 30% drop in anomaly detection accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000535,"raw_usage":{"total_tokens":2621,"prompt_tokens":1047,"completion_tokens":1574,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":663,"completion_tokens_details":{"reasoning_tokens":1492}},"tokens_in":663,"tokens_out":1574,"duration_ms":10954,"temperature":1.0,"reasoning_tokens":1492,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:51:34.803404+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Independently re-annotate a random subset of CXR-AD defect images with several trained annotators and compare the new masks against the published ones with IoU and boundary distance; also have an expert verify the geometric-consistency check on the original images. If inter-annotator agreement is low or the automated check misses misaligned contours, the pixel-level benchmark numbers should not be taken at face value.","supporting_citations":[{"cited_title":"MVTec AD–A comprehensive real-world dataset for unsupervised anomaly detection,","cited_arxiv_id":null,"evidence_quote":"Supplies the MVTec AD benchmark whose average image/pixel AUROC scores are the reference for the reported 29.78% degradation."},{"cited_title":"Spot-the- difference self-supervised pre-training for anomaly detection and segmentation,","cited_arxiv_id":null,"evidence_quote":"Provides the VisA surface-defect dataset used to train the AdaCLIP zero-shot baseline and represents the existing surface-oriented dataset regime."},{"cited_title":"Deep learning- based defect detection of metal parts: evaluating current methods in complex conditions,","cited_arxiv_id":null,"evidence_quote":"Shows the metal-surface defect focus of earlier industrial anomaly detection datasets, supporting the gap claim."},{"cited_title":"Towards total recall in industrial anomaly detection,","cited_arxiv_id":null,"evidence_quote":"Defines the feature-embedding baseline PatchCore; its CXR-AD image AUROC drops to 56.58 from 99.00 on MVTec AD."},{"cited_title":"Learning to Detect Multi-class Anomalies with Just One Normal Image Prompt,","cited_arxiv_id":null,"evidence_quote":"Defines the reconstruction-based baseline ONENIP, the best performer on CXR-AD (85.34/93.66) yet still below its MVTec result."},{"cited_title":"Adaclip: Adapting clip with hybrid learnable prompts for zero-shot anomaly detection,","cited_arxiv_id":null,"evidence_quote":"Defines the zero-shot multimodal baseline AdaCLIP; its pixel AUROC falls to 70.80 on CXR-AD, showing the localization challenge."}],"review_version":1}