{"id":"918be301-eb22-496c-8cb6-6c9d1a8d05ad","arxiv_id":"1908.01104","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"ADN, an artifact disentanglement network, reduces CT metal artifacts using only unpaired artifact-free and artifact-affected images, matching supervised methods on synthesized data.","lead":"The paper presents an unsupervised neural network that removes metal streaks from CT scans without needing paired clean and artifact images. If the clinical results hold, this could make metal artifact reduction practical for real scans where paired training data is impossible to obtain.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Paper's own caveat admits artifact code may transfer lesion-like anatomy; clinical 'generalization' is qualitative only and does not establish that ADN preserves anatomy.","rationale":"The paper has real support: the SYN experiments are quantitative, the ablations show each loss term contributes, and the source code is publicly linked, which aids reproducibility. The reader's conditional verdict is based on the same core weakness I identified: the artifact consistency loss can allow anatomy to be treated as artifact. I find additional, in-scope evidence for this concern in Sec. IV.G, where the authors explicitly state that an anatomical structure or lesion that looks like metal artifacts might also be transferred. This is not an external physics objection alone; it is the authors' own limitation statement, and it undercuts the claim that the clinical results demonstrate anatomical preservation. A concrete paired-data test with synthetic lesion-like structures, run with the released code, could settle whether the artifact code truly separates artifacts from anatomy. Given the lack of such a test, the conditional verdict remains appropriate: the method is promising and reproducible, but the most important clinical claim is not yet established.","tokens_in":12589,"tokens_out":4108,"duration_ms":48422,"concrete_test":"Use the released ADN code and the SYN dataset (or regenerate pairs with CNNMAR's simulator). Construct a held-out test set by taking artifact-free CT images and inserting synthetic lesion-like structures—low-contrast elongated dark regions oriented like streak artifacts—into tissue near a simulated implant, then synthesizing metal artifacts from the modified images. Run the trained ADN and compute PSNR/SSIM restricted to the lesion region against ground truth, plus a structural-similarity score limited to that region. Also run artifact transfer: given a lesion-containing artifact-affected image xa, transfer its artifact code onto an artifact-free image y with no lesion, and check whether the lesion appears in ŷa. If the lesion is removed or transferred at a rate comparable to true artifacts, then Eq.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that ADN is an unsupervised MAR method that removes artifacts while preserving anatomy, and therefore generalizes better than supervised models on clinical data. The least secure assumption is built into Eq. (8): the artifact consistency loss enforces (xa - x̂) ≈ (ŷa - y), a pixel-residual equality between two different anatomies. This requires the artifact code a extracted from xa to encode only the artifact, but nothing in the loss distinguishes a metal artifact from a dark shading pattern or streak-like anatomical structure. The paper itself concedes this in Sec. IV.G: 'If there is an anatomical structure or lesion that looks like metal artifacts, it might also be transferred.' Because the CL1 and CL2 evaluations are only qualitative—no ground truth, no patient-level split, no error bars—the reported removal of 'dark shadings and streaks' cannot be separated from possible removal of anatomical structure. For a clinical MAR method, erasing a lesion that resembles an artifact is exactly the failure that matters. The clinical generalization claim is therefore not yet established; it requires a quantitative content-preservation check.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ADN, an unsupervised metal artifact reduction (MAR) method for CT images. The method is based on disentangling an artifact code and a content code in latent space, using specialized encoders/decoders and a combination of adversarial, reconstruction, artifact consistency, and self-reduction losses to avoid the need for paired training data. The authors evaluate ADN on one synthesized dataset (SYN) with quantitative PSNR/SSIM and on two clinical datasets (CL1, CL2) with qualitative comparisons. They report that ADN outperforms existing unsupervised baselines on SYN, attains performance comparable to supervised models, and generalizes better to clinical data than supervised models. The source code is released publicly.","tokens_in":12782,"tokens_out":3239,"duration_ms":34852,"significance":"If the claims hold, this would be the first unsupervised MAR method that avoids paired synthetic training and may generalize better to clinical data. The architecture is well motivated, and the ablation study (M1-M4) shows that each loss component contributes meaningfully to the quantitative results on SYN. The public implementation is a strength for reproducibility. The central limitation is that the key clinical-generalization claim rests on qualitative evaluation alone, and the artifact consistency loss embodies a potentially unjustified additivity assumption that could cause loss of anatomical content.","major_comments":[{"comment":"The abstract and Sec. IV-E claim that ADN shows 'better generalization ability over the supervised models' on clinical datasets. This claim is supported only by qualitative visual comparisons; there is no quantitative metric, reader study, blinded evaluation, or statistical test on CL1 or CL2. For a clinical MAR method, the more important claim is that the method removes artifacts without erasing anatomical structures. Some form of quantitative content-preservation check (e.g., lesion/landmark preservation, or comparison of anatomical contours against a reference such as NMAR) is needed to support the generalization claim. Without it, the abstract's central claim is overstated.","section":"Sec. IV-E and abstract"},{"comment":"The artifact consistency loss Lart = E[||(xa - xhat) - (yhat_a - y)||_1] enforces that the residual removed from an artifact-affected image equals the residual added to an artifact-free image. This assumes that the artifact component is additive and independent of the underlying anatomy. The physical mechanisms of metal artifacts (beam hardening, scatter, partial volume) cited in Sec. I do not guarantee such additivity. The paper's own caveat in Sec. IV-G, that an anatomical structure or lesion that looks like metal artifacts 'might also be transferred,' indicates that the artifact code can contain content-dependent information. This undermines the claim in Sec. IV-E that ADN removes dark shadings and streaks 'without introducing secondary artifacts' while preserving anatomy. The authors should either provide evidence that the artifact code does not carry anatomy-specific content (e.g., an experiment on paired data with inserted lesions) or explicitly discuss this as a limitation and potential failure mode.","section":"Eq. (8), Sec. III-B"},{"comment":"The quantitative evaluation on SYN (Table II) is sound, but the clinical evaluation is purely qualitative without criteria, statistical analysis, or error bars. To justify the paper's contribution, the authors should provide a more rigorous clinical evaluation, for example a reader study with multiple evaluators and inter-rater agreement, or at least a quantitative proxy such as noise/streak metrics and anatomical-structure preservation scores. Alternatively, the claims in the abstract and Sec. I should be tempered to state that ADN shows promising qualitative generalization.","section":"Sec. IV-E, Figs. 7-8"}],"minor_comments":[{"comment":"Please ensure that the abstract's claim of 'better generalization ability' is consistent with the evaluation section, which itself states that only qualitative comparisons are performed on clinical data.","section":"Sec. IV-E"},{"comment":"The MUNIT result (SSIM 7.5) is far lower than the other baselines; a brief note explaining that MUNIT prioritizes diverse outputs over structural fidelity would help the reader interpret this outlying value.","section":"Table II"},{"comment":"The notation for the artifact-affected encoder E_Ia as a set of two encoders (Ec_Ia and Ea_Ia) could be visually clarified in the text; the current superscript/subscript placement in the PDF is easy to misread.","section":"Sec. III-A"},{"comment":"References [18] and [24] both refer to CycleGAN, and [19] and [25] both refer to MUNIT; these duplicates should be consolidated.","section":"References"},{"comment":"Minor grammatical issue: 'the padding of all the convolutional layers in the blocks ... are reflection padding' should be 'is reflection padding.'","section":"Sec. III-C"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a clinically important problem and the core unsupervised formulation is novel, but the most prominent claim (clinical generalization) currently exceeds the evidence provided. The artifact consistency loss has a plausible failure mode that the authors acknowledge in their own limitation note. I believe these concerns are addressable within the manuscript's scope, but they require additional experiments or substantial re-scoping of the claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe short version: this is a solid within-subfield contribution, but the headline claim—better generalization to clinical scans than supervised models—is supported only by qualitative images. The paper's own caveat in Sec. IV-G is more damaging than the authors seem to realize.\n\nWhat's new and good: the arXiv version extends the MICCAI 2019 preliminary paper with an ablation study, artifact transfer experiments, and additional baseline comparisons. The SYN results are the strongest part: ADN clearly beats the other unsupervised baselines, and Table III shows each loss term contributes. That is reproducible evidence, and the code link helps. Credit where earned: the disentanglement idea is cleanly formulated and the experiments on synthesized data are honestly reported.\n\nThe soft spots: (1) The clinical generalization claim (Sec. IV-E) has no quantitative metric, reader study, or patient-level split—only a few images. That is a real gap for a paper whose abstract promises better clinical generalization. (2) The stress-test note is on target. Eq. (8) enforces (xa - x̂) ≈ (ŷa - y), which assumes metal artifacts are additive and independent of anatomy. Beam hardening, scatter, and partial volume—which the introduction cites—are not additive. The paper's own Sec. IV-G admits that an anatomical structure or lesion that looks like metal artifacts might be transferred. For a clinical MAR method, erasing a lesion is exactly the failure that matters. The authors flag it in the synthesis context, but the same mechanism applies to reduction. A content-preservation check on paired clinical data or a reader study is needed before the central claim is credible. (3) Minor: the architecture and losses appeared in [22], so novelty is partly discounted; the CL2 dataset is not described in enough detail to reproduce; the code link has no commit hash.\n\nWho this is for: researchers in MAR or unsupervised medical image translation. It deserves a serious referee—the SYN evidence and ablation justify referee time—but the clinical claim needs major additional support. I would send it out, with the expectation of heavy revision or a resubmission.\n\nRegards.","headline":"A credible unsupervised MAR method with a clear disentanglement formulation, but the clinical generalization claim rests on qualitative images and an additivity assumption the paper itself concedes may transfer lesion-like anatomy.","tokens_in":13302,"tokens_out":1844,"would_cite":false,"duration_ms":21003,"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":"First unsupervised CT metal artifact reduction matches supervised models on synthetic data and outperforms them on clinical scans.","keywords":["metal artifact reduction","unsupervised learning","artifact disentanglement","computed tomography","image-to-image translation","generative adversarial networks","latent space disentanglement","CT artifacts"],"falsifier":"Take a set of clinical metal-affected CT images and their metal-free counterparts reconstructed from the same patient (e.g., before and after implant removal); run ADN trained only on unpaired images and check in metal-free tissue whether the difference between the network's output and the true clean image is near zero, and if the artifact code from one anatomy reproduces the true artifact when applied to a different anatomy.","tokens_in":12390,"feed_emoji":"🩻","tokens_out":4799,"duration_ms":46634,"temperature":0.7,"pith_summary":"CT scans of patients with metal implants are marred by streaking and shading artifacts that obscure anatomy. Existing deep-learning fixes train on paired images synthesized to mimic the artifacts, but synthetic artifacts do not match real clinical physics, so the models transfer poorly. This paper claims the first unsupervised alternative: a network that separates each image into a content code and an artifact code in latent space, using only unlabeled groups of artifact-affected and artifact-free CT images. With artifact reduction, transfer, and self-reconstruction losses, it removes streaks without paired data. On a synthesized test set it roughly matches supervised baselines, and on two clinical datasets it generalizes better than the supervised models it is compared with.","feed_headline":"CT metal artifacts removed without paired training images","feed_subtitle":"Unsupervised disentanglement matches supervised models on synthetic scans and generalizes better to real clinical ones.","key_machinery":"The load-bearing object is the artifact consistency loss, Eq. (8): $\\mathcal{L}_{\\mathrm{art}} = \\mathbb{E}[\\|(x_a - \\hat{x}) - (\\hat{y}_a - y)\\|_1]$, which forces the artifact subtracted from an artifact-affected image to equal the artifact added to an artifact-free image under the same artifact code. This is what lets the network learn anatomical precision without paired images. It is supported by an artifact pyramid decoding design that merges the artifact code into the content decoder at multiple scales, and by the full set of competing losses (adversarial, reconstruction, self-reduction) that keep the separated codes meaningful.","core_discovery":"The paper's central claim is that metal artifact reduction can be learned without any anatomically paired artifact-free ground truth by treating it as artifact disentanglement. The artifact disentanglement network (ADN) maps an artifact-affected image through a content encoder and an artifact encoder into separate latent spaces, and an artifact-free image through its own content encoder. Artifact removal is decoding the content code alone; artifact synthesis is decoding a content code together with an artifact code. Four losses drive the separation: adversarial losses make each output look like its target domain, a reconstruction loss keeps the autoencoding faithful, an artifact consistency loss requires the artifact stripped from an artifact image to equal the artifact added to a clean image, and a self-reduction loss requires that removing the synthesized artifact from a transferred image recovers the clean original. The paper reports that on synthesized data ADN is on par with supervised methods, while on clinical scans—where supervised models trained on synthetic artifacts degrade—ADN removes streaks and shading without introducing secondary artifacts.","pith_inferences":["The artifact consistency loss is the likeliest failure point: if real metal artifacts are non-additive and anatomy-dependent (beam hardening, scatter), the equality in Eq. (8) can be satisfied by transferring anatomy-dependent residue into the artifact code, letting the network delete real structures.","A testable extension: apply ADN to a dataset with known non-linear artifacts and compare the artifact code's transferability across anatomies; if the code carries anatomy-specific content, artifact transfer quality will drop when source and target anatomies differ greatly.","Another extension: measure whether the model's clinical generalization comes from the unsupervised objective or from the absence of synthetic-data bias; training a supervised model directly on clinical artifact-affected images with noisy pseudo-labels could isolate the cause.","If artifact synthesis is realistic enough, it could serve as data augmentation for supervised models, which would be a practical bridge between the two paradigms."],"forward_implications":["Clinics could reduce metal artifacts without needing paired scans or access to raw projection data, since ADN works directly on reconstructed CT images.","Supervised MAR models trained on synthetic artifacts fail under domain shift; an unsupervised model trained on the target domain's own unpaired images is a more robust route, as demonstrated by the clinical results.","The same architecture should transfer to other artifact types (noise, streaks, blur) because the formulation does not assume a specific artifact mechanism.","Artifact synthesis from one patient applied to another's artifact-free image could supply realistic metal-affected training data for downstream tasks like segmentation.","The self-reduction loop provides a general recipe for adding self-supervision to image translation when paired data is unavailable."],"supporting_citations":[{"why":"Supplies the metal artifact synthesis procedure used to build the SYN dataset and serves as a supervised projection-domain baseline.","marker":"[4]"},{"why":"Supervised GAN-based MAR baseline trained on synthesized pairs and compared on all datasets.","marker":"[5]"},{"why":"Backend implementation used for cGANMAR, defining the supervised Pix2Pix-style comparison.","marker":"[13]"},{"why":"CycleGAN baseline for unsupervised image-to-image translation; the paper argues its cycle consistency keeps artifacts.","marker":"[18]"},{"why":"MUNIT baseline for multimodal unsupervised translation, compared as a natural-image method unsuited to artifact reduction.","marker":"[19]"},{"why":"DRIT baseline for disentangled unsupervised translation, compared on all datasets.","marker":"[20]"},{"why":"Deep image prior baseline, used as an unsupervised artifact-reduction method for less structured artifacts.","marker":"[21]"},{"why":"DeepLesion, source of artifact-free CT images for the SYN dataset.","marker":"[31]"},{"why":"NMAR conventional projection-domain baseline with prior-image normalization.","marker":"[3]"},{"why":"Linear-interpolation projection completion baseline for conventional MAR.","marker":"[2]"}],"fun_headline_variants":["Unsupervised disentanglement strips CT metal artifacts","No paired images: unsupervised metal artifact reduction","ADN: first unsupervised CT metal artifact reduction","Metal artifacts removed without ground truth pairs","Unsupervised method generalizes better on clinical CT"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that the metal artifact in an image is an additive component that can be lifted out and re-attached to another image unchanged; if real artifacts depend on the underlying anatomy, the network may erase anatomical structures instead of artifacts.","fun_headline_variants_meta":{"raw":{"variants":["Unsupervised disentanglement strips CT metal artifacts","No paired images: unsupervised metal artifact reduction","ADN: first unsupervised CT metal artifact reduction","Metal artifacts removed without ground truth pairs","Unsupervised method generalizes better on clinical CT"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000206,"raw_usage":{"total_tokens":1392,"prompt_tokens":934,"completion_tokens":458,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":550,"completion_tokens_details":{"reasoning_tokens":389}},"tokens_in":550,"tokens_out":458,"duration_ms":7147,"temperature":1.0,"reasoning_tokens":389,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:23:33.178344+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a set of clinical metal-affected CT images and their metal-free counterparts reconstructed from the same patient (e.g., before and after implant removal); run ADN trained only on unpaired images and check in metal-free tissue whether the difference between the network's output and the true clean image is near zero, and if the artifact code from one anatomy reproduces the true artifact when applied to a different anatomy.","supporting_citations":[{"cited_title":"Convolutional neural network based metal artifact reduction in x-ray computed tomography,","cited_arxiv_id":null,"evidence_quote":"Supplies the metal artifact synthesis procedure used to build the SYN dataset and serves as a supervised projection-domain baseline."},{"cited_title":"Conditional genera- tive adversarial networks for metal artifact reduction in ct images of the ear,","cited_arxiv_id":null,"evidence_quote":"Supervised GAN-based MAR baseline trained on synthesized pairs and compared on all datasets."},{"cited_title":"Unpaired image-to-image translation using cycle-consistent adversarial networks,","cited_arxiv_id":null,"evidence_quote":"CycleGAN baseline for unsupervised image-to-image translation; the paper argues its cycle consistency keeps artifacts."},{"cited_title":"Multimodal unsuper- vised image-to-image translation,","cited_arxiv_id":null,"evidence_quote":"MUNIT baseline for multimodal unsupervised translation, compared as a natural-image method unsuited to artifact reduction."},{"cited_title":"Diverse image-to- image translation via disentangled representations,","cited_arxiv_id":null,"evidence_quote":"DRIT baseline for disentangled unsupervised translation, compared on all datasets."},{"cited_title":"Deep image prior,","cited_arxiv_id":null,"evidence_quote":"Deep image prior baseline, used as an unsupervised artifact-reduction method for less structured artifacts."},{"cited_title":"Deeplesion: automated mining of large-scale lesion annotations and universal lesion detection with deep learning,","cited_arxiv_id":null,"evidence_quote":"DeepLesion, source of artifact-free CT images for the SYN dataset."},{"cited_title":"Normalized metal artifact reduction (nmar) in computed tomography,","cited_arxiv_id":null,"evidence_quote":"NMAR conventional projection-domain baseline with prior-image normalization."},{"cited_title":"Reduction of ct artifacts caused by metallic implants","cited_arxiv_id":null,"evidence_quote":"Linear-interpolation projection completion baseline for conventional MAR."}],"review_version":1}