{"id":"f94b5803-01c1-41a5-9dcd-a32040aef834","arxiv_id":"2505.04105","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"MAISY improves simulated CT motion correction by combining SAM segmentation masks with an attention GAN and a variance-selective SSIM loss, though reported headline gains are versus a weaker baseline and no error bars or code are provided.","lead":"A CT motion-correction model called MAISY uses a segmentation model to focus corrections on organ borders and a variance-weighted SSIM loss to preserve fine detail. Tests on simulated chest and head CT scans beat earlier GAN methods, but the abstract's large 40% gain is from a weaker baseline, not the true state of the art.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Phantom-only evaluation with no real patient motion data leaves the clinical superiority claim unsupported.","rationale":"I read the paper as making an empirical performance claim: MAISY outperforms state-of-the-art motion correction on CT. The necessary conditions are a representative test distribution and a fair baseline. The weaker of these is representativeness: the entire evaluation uses XCAT phantoms with simulated motion (Section 3.1), so the test distribution is a simulation generated by the same pipeline that defines the task. The reader's weakest_assumption identifies this same issue, and I agree. The abstract's 40/10/16 percentages are computed against cGAN rather than against the best baseline AGGAN, which is a reporting and interpretation problem, but it does not change the core concern: even the honest head-to-head margins over AGGAN are only demonstrated in simulation. The ablation and modularity studies are internally consistent and directionally supportive, so I would not reject the method; the uncertainty is about transfer, not about internal consistency. A real-data benchmark with error bars is the one check that would resolve whether the claimed advantage survives outside the simulator. For these reasons I keep the reader's CONDITIONAL verdict unchanged.","tokens_in":6761,"tokens_out":6196,"duration_ms":66279,"concrete_test":"Obtain a benchmark of real patient chest CT scans with motion-corrupted acquisitions and motion-free reference scans (or expert-validated artifact labels). Train MAISY and AGGAN on identical patient-disjoint splits, using 5 random seeds, and report the paired mean difference and 95% bootstrap CI for PSNR/SSIM/Dice on held-out patients. If the CI for MAISY−AGGAN includes 0 or is negative, the phantom-only superiority claim does not transfer to clinical data.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim requires that the simulated XCAT motion artifacts described in Section 3.1 are representative of real patient motion during CT acquisition. This condition is not secured: all training and test images are generated from XCAT phantoms with computer-simulated rigid and non-rigid motion, and no real patient CT volume with motion artifacts is used. Section 5 lists only 'other modalities' as a limitation, not the absence of real data. If real motion includes patterns absent from the simulator—non-rigid organ drift, discontinuous breathing, involuntary twitches, or interactions with metal and beam hardening—the reported gains over AGGAN (3.20% PSNR, 0.71% SSIM, 0.81% Dice on chest; Table 1) may not transfer. The abstract amplifies the effect by quoting 40%/10%/16% improvements relative to cGAN rather than to the best baseline, but the load-bearing gap is the lack of any real acquisition evidence.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes MAISY, a GAN-based framework for correcting patient motion artifacts in CT images. The method combines a pre-trained Segment Anything Model (SAM) to produce anatomical mask features that are concatenated with the input to an Attention-Guided GAN (AGGAN), and a new Variance-Selective SSIM (VS-SSIM) loss that averages SSIM over the highest-variance patches. Experiments on simulated XCAT chest and head/jaw CT datasets are compared with U-Net, cGAN, Pix2Pix, Cycle GAN, and AGGAN, and the authors report improvements in PSNR, SSIM, and Dice, along with ablation and modularity studies. The abstract claims a 40% PSNR increase, 10% SSIM increase, and 16% Dice increase, which turn out to be relative to cGAN, not the strongest baseline.","tokens_in":6931,"tokens_out":6836,"duration_ms":65419,"significance":"If the reported results are reproducible, the paper makes a useful engineering contribution: adding SAM-derived structural guidance and a variance-weighted patch SSIM loss to an existing GAN architecture yields consistent improvements on the two simulated datasets and also improves the performance of other generators when grafted onto them. The modularity study is a strength, as it demonstrates that the proposed modules are transferable. However, the clinical significance of the work is not established because all experiments are based on simulated phantom data, and the headline improvement figures in the abstract are measured against a weak baseline. The evaluation also lacks error bars or statistical tests, making it difficult to assess whether the claimed margins are meaningful.","major_comments":[{"comment":"The abstract states that PSNR increases by 40%, SSIM by 10%, and Dice by 16%, but these figures are relative to the cGAN baseline (PSNR 31.210, SSIM 0.901, Dice 0.855 on chest). In contrast, the improvement over the strongest baseline AGGAN is only 3.20% PSNR, 0.71% SSIM, and 0.81% Dice on the chest dataset. Since the paper's central claim is that MAISY outperforms state-of-the-art counterparts, the comparison should be against the best-performing baseline, not a weaker one. The abstract overstates the gains and should be revised to reference the correct baseline.","section":"Abstract and Table 1"},{"comment":"All experiments use simulated XCAT phantom data with computer-simulated rigid and non-rigid motion. No real patient CT volumes with motion artifacts are used, and Section 5 only notes that other modalities were not tested; it does not acknowledge the absence of real acquisition data. If real patient motion includes patterns absent from the simulator (e.g., discontinuous breathing, organ drift, or interaction with metal and beam hardening), the reported gains over AGGAN may not transfer to clinical scans. The authors should either include a real-data validation or substantially temper the claims of clinical applicability.","section":"Section 3.1 and Section 5"},{"comment":"The definition of VS-SSIM is ambiguous and irreproducible as written. The equation contains an unexplained prefactor 3, while the text says the SSIM values are 'averaged' over Ω patches, which would correspond to 1/|Ω|. If the factor 3 is a typo, it should be corrected; if intentional, its purpose and effect on the loss bounds must be explained. Additionally, the number of top-variance patches Ω, the patch size, and the loss balancing weights λ1 and λ2 are never specified anywhere in the paper. Without these hyperparameters, the method cannot be independently reproduced.","section":"Section 2.2, Eq. (5)"},{"comment":"The training data setup is unclear. Equation (1) and the VS-SSIM loss use the ground truth y_gt, which implies paired training data. However, the text also says that AGGAN compares the generated output against 'an unpaired motion-free target image' and uses a cycle loss. It is not clear whether the authors used paired, unpaired, or semi-supervised training. This distinction is essential for interpreting the loss function and for understanding why a cycle consistency loss is included when ground truth images appear to be available. The authors should clarify the training paradigm.","section":"Section 2.2 and Eq. (1)"},{"comment":"No error bars, confidence intervals, or significance tests are reported for any of the quantitative results. The head-and-jaw test set consists of only 20 volumes, so the differences between methods (e.g., Dice 0.998 vs. 0.999) may not be statistically meaningful. The authors should report standard deviations across repeated runs or bootstrap confidence intervals to support the claim that the improvements are consistent and not simply due to random variation.","section":"Tables 1, 3, and 4"}],"minor_comments":[{"comment":"There are typographical errors: 'bowl' should be 'bowel' (involuntary motion examples), and 'Patient motions encompasses' should be 'Patient motion encompasses'. Also, 'sematic' in Section 2.1 should be 'semantic'.","section":"Introduction"},{"comment":"The Dice score computation is incompletely described: the paper states that a threshold is applied to create a binary mask, but the threshold value is not provided. Please specify the thresholding procedure.","section":"Section 3.2"},{"comment":"The abbreviation 'FBCT' is not defined; the text says 'step-and-shoot flat-bed CT (FBCT)' but does not explain what FBCT stands for in this context or provide a description of the geometry.","section":"Section 3.1"},{"comment":"The symbols R_gen and M_gen (content mask and attention mask) are not defined in the text. They appear to be borrowed from the original AGGAN, but the adaptation should be clarified so the reader can follow the architecture.","section":"Section 2.2, Eq. (4)"},{"comment":"The mask feature map is described as 'pre-binarized', but the binarization threshold is not given. Also, the paper does not state whether the SAM encoder is frozen or fine-tuned during training, which affects the interpretation of the 'dynamic learning' claim.","section":"Section 2.1"},{"comment":"The table numbering skips Table 2: the manuscript jumps from Table 1 in Section 4.1 to Table 3 in Section 4.2. Please renumber or add the missing table.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The manuscript describes a plausible and generally sound engineering approach, and the modularity study is a positive feature. However, the abstract overstates the improvements by comparing to the weakest baseline rather than the strongest, and the evaluation is confined to simulated phantom data with no real patient validation. The lack of stated hyperparameters and the ambiguous loss definition also impede reproducibility. I believe the paper can be made publishable with a thorough revision that includes either a real-data experiment or a much more cautious presentation of the claims, proper statistical reporting, and a complete specification of the method's parameters."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nTwo things to know about arXiv:2505.04105. First, it is a legitimate incremental contribution: combining SAM-generated boundary masks with an attention GAN and a new variance-selective SSIM loss for CT motion correction. The ablations and modularity study are the best part—they show each component adds value across three baseline architectures, and the method fine-tunes to a new anatomical region (head/jaw) with only 20 images, which is a real plus. Second, the abstract's 40%/10%/16% gains are against the weakest baseline (cGAN), not against the state of the art (AGGAN); the actual margins over AGGAN are 3.2% PSNR, 0.7% SSIM, 0.8% Dice on chest. That is a meaningful difference but not the story the abstract tells.\n\nThe method itself is sensible. The VS-SSIM loss that selects the top-variance patches is a reasonable patch-level extension of SSIM, and using SAM as a prior to focus correction inside anatomical boundaries is well motivated. The paper is clearly written and the setup is reproducible in principle.\n\nThe soft spots are real and mostly cluster around the evaluation. All training and test images are XCAT phantoms with computer-simulated rigid/non-rigid motion. No real patient CT with motion artifacts is used anywhere. The limitation section only says other modalities were not tested—it omits the more important caveat that no real data was used. For a paper whose title says 'medical image motion correction' and whose abstract suggests clinical relevance, that is a serious gap. The reported numbers may not survive contact with real patient motion, which includes breathing drift, twitches, and metal/beam-hardening interactions that the simulator may not capture.\n\nAlso missing: error bars, significance tests, and several hyperparameters (loss weights, the number of top patches, SAM prompt placement). These are minor in comparison but should be addressed.\n\nThe internal comparison is fair—held-out test split, same simulated motion pipeline for all methods—so the relative ranking is credible within the simulation. The paper is best read as a proof-of-concept for combining foundation-model masks with attention GANs and adaptive losses. It is for researchers working on deep learning CT motion correction, not for radiologists seeking a clinically validated tool.\n\nYes, this deserves peer review. It is a solid engineering paper with a clear experimental design. A good reviewer should ask for real-data validation, error bars, and an honest abstract before acceptance; those are reasonable major revisions.","headline":"Solid incremental CT motion-correction paper with honest ablations, but the abstract overstates gains and the phantom-only evaluation needs real-data validation before the clinical claims can be taken at face value.","tokens_in":7452,"tokens_out":2239,"would_cite":true,"duration_ms":21476,"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":"MAISY claims CT motion correction improves sharply when an attention-guided GAN is steered by SAM anatomical masks and a variance-selective SSIM loss that focuses on high-variance patches.","keywords":["motion correction","CT","GAN","Segment Anything Model","SSIM loss","variance-selective loss","XCAT phantom","motion artifacts"],"falsifier":"Collect real clinical CT exams with motion artifacts and expert-validated artifact-free reconstructions, run MAISY and the strongest baseline AGGAN on them, and compare PSNR, SSIM, Dice, and radiologist readings; if MAISY's margins shrink to near zero or reverse, the paper's central claim fails.","tokens_in":6555,"feed_emoji":"🩻","tokens_out":8680,"duration_ms":79935,"temperature":0.7,"pith_summary":"MAISY is a generative method for correcting patient motion artifacts in CT images. It claims that motion correction improves when the network is explicitly told where artifacts concentrate: a Segment Anything Model (SAM) mask marks anatomical boundaries, and a Variance-Selective SSIM (VS-SSIM) loss focuses training on high-variance patches, the regions most affected by motion. On simulated chest and head CT XCAT phantom data, the paper reports the highest PSNR, SSIM, and Dice among all compared methods; against the second-best method, AGGAN, MAISY gains 3.20% PSNR on the chest dataset and 13.91% PSNR on the head-and-jaw dataset. The significance is that motion artifacts, which blur organs and obscure pathology, can be removed without globally smoothing away diagnostically useful local detail.","feed_headline":"CT motion correction gains 3.2% PSNR over the best prior GAN","feed_subtitle":"SAM boundary masks plus variance-selective SSIM preserve anatomy while removing blur and ghosting in chest and head CT.","key_machinery":"The argument is carried by two complementary modules. The first is SAM-based masking: the Segment Anything Model produces a pre-binarized mask feature map that semantically marks anatomical boundaries, acting as a pre-attention guide for the AGGAN's attention mask. The second is the VS-SSIM loss, a patch-level structural similarity loss that selects the top-variance patches $\\Omega$, computes SSIM on each patch, and averages them. Together these modules concentrate the generator's capacity on the anatomical regions where motion artifacts are strongest, while the adversarial and cycle losses preserve global structure.","core_discovery":"MAISY's central claim is that CT motion correction benefits from combining anatomical localization with a variance-weighted structural loss. The method computes a mask feature map $F = D(E(y_{input}), P)$ from SAM using point prompts, concatenates it with the motion-corrupted image, and feeds the result into an Attention Guided GAN (AGGAN) whose attention mask $M$ highlights regions needing correction. The VS-SSIM loss ranks local patches by pixel variance and averages SSIM over the top-variance patches $\\Omega$, giving $L_{VS\\text{-}SSIM} = 1 - \\frac{1}{|\\Omega|}\\sum_{i\\in\\Omega}\\mathrm{SSIM}(y_{gt,i}, y_{gen,i})$; the total loss combines this with adversarial and cycle-consistency terms. On chest CT, MAISY reports PSNR 43.573, SSIM 0.989, and Dice 0.991 versus AGGAN's 42.211, 0.982, and 0.983, and on the head-and-jaw dataset PSNR 43.372 versus 38.075, with the gains attributed to focusing correction along anatomical boundaries instead of applying global smoothing.","pith_inferences":["Beyond the paper: the headline gains of 40% PSNR, 10% SSIM, and 16% Dice in the abstract are computed against the weak cGAN baseline; compared with the closest competitor, AGGAN, the chest-dataset gains are 3.20% PSNR, 0.71% SSIM, and 0.81% Dice.","Beyond the paper: all results come from simulated XCAT phantom scans with synthetic rigid and non-rigid motion, so unmeasured clinical factors such as breathing, cardiac motion, and peristalsis could narrow the margins in real patient data.","Beyond the paper: the SAM prompts are manual point prompts, so a clinical deployment would need automatic prompt selection; the paper does not specify how that would work.","Beyond the paper: the variance-selective loss idea is not tied to the GAN backbone and could be tested in diffusion-based correction models to see whether the loss or the architecture drives the improvement."],"forward_implications":["MAISY's modules are portable: the paper's modularity study shows that adding SAM masking and VS-SSIM to cGAN, Pix2Pix, and Cycle GAN raises their PSNR, SSIM, and Dice.","The method transfers to a new anatomy with little retraining: after 10 epochs of fine-tuning on 20 head-and-jaw scans, MAISY stays accurate while the comparison methods drop in performance when fine-tuned on a different anatomy.","Dice scores near 0.99 indicate that corrected images preserve organ boundaries, so motion correction does not have to sacrifice structural fidelity.","Concentrating the loss on high-variance patches should protect small, diagnostically relevant details from being smoothed away.","If the XCAT simulation faithfully models clinical motion, the reported margins should translate directly into better artifact removal on real chest and head CT acquisitions."],"supporting_citations":[{"why":"Supplies the Attention Guided GAN backbone that MAISY adapts, including the attention mask and generator architecture.","marker":"[14]"},{"why":"Supplies the Segment Anything Model whose mask feature map localizes anatomical boundaries in the corrected image.","marker":"[13]"},{"why":"Supplies the XCAT phantom chest and head CT datasets with simulated rigid and non-rigid motion used in all experiments.","marker":"[15]"},{"why":"Documents SSIM's limitations with varying pixel intensity, luminance, and variance, which motivates the VS-SSIM loss.","marker":"[12]"},{"why":"Provides the cGAN baseline used in the comparison and as a target in the modularity study.","marker":"[5]"},{"why":"Provides the Pix2Pix baseline used in the comparison and modularity study.","marker":"[16]"}],"fun_headline_variants":["SAM boundary masks plus variance-selective SSIM lift CT PSNR 3.2%","Anatomy-aware loss sharpens motion-corrupted CT, beats GANs","MAISY uses SAM masks to fix CT motion blur, gains 3.2% PSNR","Variance-selective SSIM preserves details in CT motion correction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that motion artifacts created by computer simulation on XCAT phantom CT images behave like motion artifacts in real patient CT scans, so phantom-measured improvements will transfer to clinical use.","fun_headline_variants_meta":{"raw":{"variants":["SAM boundary masks plus variance-selective SSIM lift CT PSNR 3.2%","Anatomy-aware loss sharpens motion-corrupted CT, beats GANs","MAISY uses SAM masks to fix CT motion blur, gains 3.2% PSNR","Variance-selective SSIM preserves details in CT motion correction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000666,"raw_usage":{"total_tokens":3088,"prompt_tokens":1044,"completion_tokens":2044,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":660,"completion_tokens_details":{"reasoning_tokens":1954}},"tokens_in":660,"tokens_out":2044,"duration_ms":13242,"temperature":1.0,"reasoning_tokens":1954,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:36:45.535587+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Collect real clinical CT exams with motion artifacts and expert-validated artifact-free reconstructions, run MAISY and the strongest baseline AGGAN on them, and compare PSNR, SSIM, Dice, and radiologist readings; if MAISY's margins shrink to near zero or reverse, the paper's central claim fails.","supporting_citations":[{"cited_title":"Attentiongan: Unpaired image-to-image translation using attention-guided generative adversarial networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the Attention Guided GAN backbone that MAISY adapts, including the attention mask and generator architecture."},{"cited_title":"Segment anything,","cited_arxiv_id":null,"evidence_quote":"Supplies the Segment Anything Model whose mask feature map localizes anatomical boundaries in the corrected image."},{"cited_title":"Evaluation of motion artefact reduction depending on the artefacts’ directions in head MRI using conditional generative adversarial networks,","cited_arxiv_id":null,"evidence_quote":"Provides the cGAN baseline used in the comparison and as a target in the modularity study."},{"cited_title":"Image-to-image translation with conditional adversarial networks,","cited_arxiv_id":null,"evidence_quote":"Provides the Pix2Pix baseline used in the comparison and modularity study."}],"review_version":1}