{"id":"b23d37b1-0ec2-4c66-9cfa-f92b77983075","arxiv_id":"2411.17488","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A whole-body MR-to-CT synthesis framework using attention gates, registration-aware training targets, and contrastive organ-semantic alignment improves synthetic CT quality and PET attenuation correction over CycleGAN baselines.","lead":"A three-module deep learning pipeline synthesizes CT-like images from whole-body MR scans for PET attenuation correction, combining structure-guided attention gating, learned registration, and contrastive organ-semantic alignment. On a 350-subject multi-center dataset it reports higher PSNR/SSIM than CycleGAN baselines and lower SUV errors in PET reconstruction.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table I comparison is confounded: proposed method is trained/evaluated against SpatAlign-aligned CT while CycleGAN baselines lack this aligned supervision; Table II's vanilla baseline already beats all CycleGAN variants, so the claimed superiority may be an artifact of access to registration.","rationale":"The reader's weakest_assumption focuses on MOOSE segmentation reliability, which is indeed a robustness concern: semantic alignment and tissue-aware registration both depend on MOOSE masks (Section III-B.1 and III-C), and the paper concedes occasional poor segmentation can compromise contrastive learning (Section V). However, I find a more load-bearing issue in the evaluation protocol itself. The paper's strongest claim is quantitative superiority over whole-body MR-to-CT synthesis baselines (Section IV-B.1, Table I), but the comparison is confounded by the asymmetric use of the SpatAlign-aligned CT. Because the proposed method is explicitly trained to match I_align while the CycleGAN baselines are not, the reported margins may reflect this difference in supervision rather than the effectiveness of the three proposed modules. Table II supports this reading: the authors' own vanilla supervised baseline (row 1) already outperforms all CycleGAN baselines by a wide margin (20.96 vs 19.16), while the full method's gain over that vanilla baseline is only 0.75 dB. This suggests that the main driver of the apparent superiority is direct supervision with a well-aligned target, not the gated U-Net, tissue-aware registration, or contrastive learning. The paper does not clarify whether Table I metrics use the SpatAlign-aligned CT or the original/B-spline-aligned CT as the reference, and this ambiguity directly affects the validity of the central claim. I therefore propose a controlled comparison that would settle the issue. My verdict remains CONDITIONAL because the concern can be addressed by additional experiments and reporting; but it is a more fundamental condition than the MOOSE robustness issue, because it bears on whether the headline result actually measures what it claims.","tokens_in":14689,"tokens_out":6843,"duration_ms":63608,"concrete_test":"Train a simple 3D U-Net with L1 loss directly on SpatAlign-aligned CT (same data, no attention gates, no contrastive loss), and additionally retrain each CycleGAN baseline with direct L1 supervision on the same SpatAlign-aligned CT. Evaluate all methods against both the Elastix-aligned CT and the SpatAlign-aligned CT, reporting PSNR/SSIM for each combination. If the full method's advantage over the simple U-Net is not significantly larger than the simple U-Net's advantage over the CycleGAN baselines, the claimed superiority is largely a consequence of access to aligned supervision rather than of the proposed structure and semantic modules.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim (Section IV-B.1: 'our proposed framework outperforms all other approaches') rests on Table I, but the comparison is not controlled for access to aligned supervision. SpatAlign (Section III-B) produces I_align that serves as the ground truth for training the synthesis network (Eq. 2) and, presumably, for computing the PSNR/SSIM values in Table I. The CycleGAN baselines are unpaired methods that do not receive this aligned target; at best they are trained and evaluated against the initial B-spline-aligned CT. Evidence of the confound is visible in Table II: the authors' own baseline, a vanilla supervised U-Net with conventional registration, reaches whole-body PSNR 20.96, already exceeding the best CycleGAN baseline (gc-CycleGAN, 19.16). The gap between the full method (21.71) and the naive baseline (20.96) is smaller than the gap between the naive baseline and the CycleGAN baselines, indicating that direct supervision from the aligned CT—not the proposed modules—is the dominant factor. If Table I uses the SpatAlign-aligned CT as reference, the comparison is unfair because the baselines never see this target. If it uses the original CT as reference, the explicit registration drives the synthetic CT away from that reference, also penalizing the proposed method. The paper never states exactly which reference is used for the reported quantitative metrics.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a whole-body MR-to-CT synthesis framework for PET attenuation correction, composed of three modules: a Structure-Guided Synthesis network with gated U-Net and attention gates, a Spatial Alignment module that registers paired CT to MR using a tissue-aware MINE-based similarity loss and a respiration-aware smoothness regularization, and a Semantic Alignment module that applies contrastive learning on MOOSE-derived organ features. The method is evaluated on a 350-subject multi-center dataset against CycleGAN variants and a segmentation-based attenuation correction approach, reporting PSNR/SSIM and SUV differences. The authors also provide an ablation study showing monotonic gains when adding each module.","tokens_in":15010,"tokens_out":6722,"duration_ms":59371,"significance":"The work addresses a clinically important problem and the three-module design is clearly motivated by two real challenges in whole-body synthesis: misalignment and complex intensity mapping. The use of a large multi-center dataset and the downstream PET attenuation correction evaluation are strengths. The ablation is internally consistent and each proposed component contributes positively in the reported experiments. However, the main comparative claim—that the framework outperforms existing methods—is currently not fully established because the evaluation reference is ambiguous, the comparison to CycleGAN baselines is not controlled for access to aligned supervision, and no statistical significance testing is provided. Given these issues, the technical contribution is promising but needs additional validation.","major_comments":[{"comment":"The paper never states whether the PSNR/SSIM values in Table I are computed against the original CT volume I^CT_in or the SpatAlign-aligned volume I^CT_align. This is load-bearing because Eq. (2) trains the synthesis network against I^CT_align, and the comparison in Table I is against unpaired CycleGAN baselines that never see this aligned target. If the evaluation uses I^CT_align as the reference, the baselines are systematically penalized by residual spatial misalignment, making the comparison unfair; if it uses I^CT_in, then the proposed method is explicitly optimized toward a target that differs from the reference, and the metric conflates synthesis quality with registration. The evaluation reference must be stated explicitly, and the main comparison should include a controlled supervised baseline that also receives the aligned supervision (e.g., the vanilla U-Net from Table II) in Table I itself, not only in the ablation.","section":"IV-B.1, Table I, Eq. (2)"},{"comment":"All quality metrics and SUV differences are reported as point estimates without per-subject standard deviations, confidence intervals, or significance tests. Since the dataset has 50 test subjects, paired comparisons are possible, and claims such as 'significantly lower standard deviations' (Section IV-C) require statistical support. Without error bars, the differences between the full method (PSNR 21.71) and the vanilla baseline (20.96) in Table II, and the differences against CycleGAN baselines in Table I, could be within subject variability. Please provide mean±SD over test subjects and pairwise tests (e.g., Wilcoxon signed-rank) for each metric and region.","section":"Tables I and III"},{"comment":"The tissue-aware weight map in SpatAlign and the positive/negative pairs in SemAlign rely on MOOSE segmentation masks applied to synthetic and aligned CT images, but the manuscript does not quantify segmentation accuracy on these images, which are 128×128×128 with 2 mm spacing and include synthetic CT data outside the training distribution of MOOSE. The Discussion (Section V) concedes that 'occasional poor segmentation outcomes can compromise the efficacy of contrastive learning,' yet no sensitivity analysis or failure statistics are reported. If MOOSE masks are unreliable on synthetic CT, both the MINE sampling weights and the InfoNCE loss in Eq. (9) are driven by noisy labels, and the advertised contribution of SemAlign is not robustly demonstrated. Please report native MOOSE accuracy on the aligned CT (e.g., Dice against manual labels on a subset) and on synthetic CT during training, or provide an ablation that degrades mask quality.","section":"III-B.1, III-C, and V"},{"comment":"The definition of the tissue weight map is not reproducible as written: 'define its initial weight as the reciprocal of its volume ratio P6 i=1 Vi / Vi' is dimensionally inconsistent and should presumably read 1/(V_i/Σ_j V_j). The authors also do not report the sensitivity of the registration to the number of sampled voxel pairs n=100,000 or to the choice of the six sub-regions. These details are needed to support the SpatAlign component.","section":"III-B.1"}],"minor_comments":[{"comment":"The caption enumerates columns (a), (b), (c), (d), (e), (f), and (h) but omits (g); please correct the enumeration to match the displayed panels.","section":"Fig. 3"},{"comment":"The phrase 'multi-layer perception' should be 'multi-layer perceptron'.","section":"Section III-B"},{"comment":"The full model is reported with spine PSNR 15.9 in Table I and 15.90 in Table II; please be consistent with decimal places.","section":"Table I and Table II"},{"comment":"The synthesis objective sums three loss terms without weighting coefficients; please specify the relative weights used in training, since the adversarial loss and reconstruction losses are on very different scales.","section":"Eq. (4)"},{"comment":"The 'SUV difference' metric is not defined; please state whether it is the mean signed difference, mean absolute difference, or a percentage difference, and specify the PET reconstruction settings.","section":"Section IV-C"},{"comment":"The Canny edge detector thresholds for computing E^CT_gt are not given; please add them for reproducibility.","section":"Section III-A"}],"recommendation":"major_revision","confidential_remarks":"The paper attacks a relevant problem and the technical design is plausible. The central concern is the evaluation protocol: the claim of superiority over CycleGAN-based methods is not established without a controlled comparison and a clear statement of the evaluation reference. I would encourage the editor to require the authors to specify the reference used for PSNR/SSIM, add a supervised baseline with aligned CT in the main comparison, and report subject-level statistics. The dependence on MOOSE also needs to be quantified. If these issues are addressed, the paper could be a valid contribution to the MR-to-CT synthesis literature."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a competent applied paper on whole-body MR-to-CT synthesis, but the headline comparison against CycleGAN is not controlled. The authors compare a paired, supervised pipeline against unpaired CycleGAN variants; their own supervised baseline (Table II) already beats every CycleGAN number. That does not sink the paper, but it means the claimed 'superiority' is mostly access to aligned supervision, not the three modules.\n\nWhat is actually new: the combination of structure-guided attention gates, a VoxelMorph-style registration with tissue-volume weighting and a respiration-aware smoothness mask, and MOOSE-based contrastive semantic alignment. Each piece is adapted from known work, but the integration for whole-body PET/MR is a legitimate extension, and the 350-subject multi-center dataset plus PET SUV validation give it some weight. The ablation is monotonic and the qualitative figures show plausible gains. Credit where due: the authors candidly state in Section V that the method needs paired data and that poor segmentation can hurt contrastive learning.\n\nSoft spots, in order of severity. First, the evaluation reference for Tables I and II is never stated. The synthesis network is trained against SpatAlign-aligned CT, so if PSNR/SSIM also use that aligned image as reference, the CycleGAN baselines are measured against a target they never saw. If the reference is the original CT, the proposed method is penalized for the very registration it uses. The paper needs to state the reference and, ideally, report both. Second, no error bars or significance tests; the gap between the full method and the vanilla supervised baseline is small (21.71 vs 20.96), so without variance we can't tell if the modules actually matter. Third, there is a minor inconsistency: Table I reports ribs PSNR 13.25 for the full method, Table II reports 13.35. Fourth, no code or data, which limits reproducibility.\n\nWho is it for: people working on MR-based attenuation correction or whole-body synthesis. It deserves serious peer review because the method is plausible and the clinical motivation is real, but it needs a major revision on the evaluation protocol. I'd want the authors to compare against a paired supervised baseline trained on the same aligned target, report metrics against both aligned and original CT references, and add error bars. Then the contribution would be clearer.","headline":"Solid applied MR-to-CT synthesis paper, but the claimed edge over CycleGAN is mostly supervised alignment, and the evaluation reference is ambiguous.","tokens_in":15508,"tokens_out":3200,"would_cite":false,"duration_ms":27161,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that whole-body MR-to-CT synthesis for PET/MR attenuation correction becomes substantially more accurate when the generator is trained with explicit spatial registration and organ-level contrastive learning alongside…","keywords":["MR-to-CT synthesis","PET/MR imaging","attenuation correction","image registration","contrastive learning","whole-body imaging","synthetic CT","structure-guided attention"],"falsifier":"Run the same region-wise PSNR, SSIM, and PET SUV-difference evaluation on an independent multi-center whole-body PET/MR dataset with paired CT: if the framework's whole-body PSNR advantage over the strongest unpaired-translation baseline disappears, or if its spine and femur SUV errors are not lower than the segmentation-based UTE method, the claimed superiority is refuted. A more targeted check is to corrupt or remove the organ-segmentation masks during training; if the framework's gain over the no-semantic-alignment ablation vanishes, the semantic-alignment claim is the operative cause.","tokens_in":14496,"feed_emoji":"🩻","tokens_out":8703,"duration_ms":77559,"temperature":0.7,"pith_summary":"Whole-body PET/MR scanners need an attenuation map for accurate PET quantification, but MR images carry no direct tissue-density information. The paper tries to establish that a synthetic CT computed from water-fat MR images can serve as that attenuation map, provided the synthesis is trained with explicit spatial and semantic alignment rather than unpaired translation alone. It proposes a three-module framework: a structure-guided synthesis network, a registration module that aligns paired CT ground truth to the MR before supervision, and a contrastive module that keeps the same organs semantically close in synthetic and real CT. On a 50-subject test set the framework reports whole-body PSNR 21.71 dB versus 19.16 dB for the best baseline, and PET SUV differences with lower mean and standard deviation in spine and femur than a segmentation-based UTE method. If these results hold, the approach offers a direct route to more reliable MR-based attenuation correction in clinical whole-body PET/MR imaging.","feed_headline":"Aligned ground truth boosts whole-body MR-to-CT synthesis by 2.5 dB","feed_subtitle":"Synthetic CT from the three-module pipeline lowers PET SUV errors in spine and femur versus standard synthesis.","key_machinery":"The load-bearing mechanism is the joint training of the synthesis network, the registration network, and the contrastive feature space. The generator is a 3D U-Net with two decoders, one for the CT image and one for its edge map, and structure-guided attention gates that gate encoder features by edge-decoder features at each scale, suppressing unwanted soft-tissue contours. The registration network estimates a deformation field between paired MR and CT, with a mutual-information neural estimator as the similarity term, weighted by inverse tissue-volume ratios, and a smoothness term that skips the thoracic-cavity boundary so rib motion is not over-smoothed. The semantic stream uses a fixed pre-trained organ segmenter to define positive and negative pairs and an InfoNCE-style contrastive loss across synthetic and aligned CT features. The key identity is that the aligned CT, not the raw paired CT, becomes the ground truth for both the reconstruction loss and the contrastive pairs, converting a spatially misaligned problem into one with explicit supervision.","core_discovery":"The central discovery is that the two obstacles to whole-body MR-to-CT synthesis—spatial misalignment between paired scans and the complex, many-tissue MR-to-CT intensity mapping—can be attacked simultaneously by making registration and semantics explicit parts of the training loop. Rather than relying only on adversarial or cycle-consistency losses, the generator is supervised by a CT volume that has been deformably registered to the MR scan; that registration is tissue-aware, sampling voxels inversely to sub-region volume, and respiration-aware, excluding the thoracic-cavity slip boundary from smoothness regularization. A second stream extracts organ-level feature vectors from a pre-trained multi-organ segmentation network and applies a contrastive loss so that matching organs in synthetic and real CT move together while different organs separate. The paper's evidence is that each added module improves PSNR and SSIM, with the full framework producing the highest scores in every reported organ region, and that PET images reconstructed with the synthetic attenuation maps show smaller SUV errors than the UTE segmentation baseline and a structure-constrained unpaired-translation baseline.","pith_inferences":["Editorial inference: the contrastive semantic alignment could be made self-supervised by using features of the aligned CT itself or by gating the contrastive pairs with segmentation confidence, which would remove the paper's stated dependence on high-quality organ masks.","Editorial inference: the same registration-plus-semantics recipe likely transfers to other misaligned cross-modality synthesis tasks, such as MR-to-dose or CBCT-to-CT, because sliding organs and organ-specific intensity mapping are not unique to whole-body PET/MR.","Editorial inference: a direct test of the semantic module would be to measure segmentation Dice between synthetic CT and real aligned CT; if synthetic organs segment poorly, the contrastive pairs are learning from noisy labels and the robustness gains would shrink as segmentation quality drops."],"forward_implications":["Whole-body MR-based attenuation correction can move from discrete tissue classes to continuous synthetic CT values, reducing the bone-related SUV underestimation seen in segmentation-based UTE methods.","Training MR-to-CT synthesis with explicitly registered paired data gives direct voxel-level supervision, anchoring synthetic images to true anatomy instead of relying only on adversarial or cycle consistency.","Adding organ-level contrastive supervision removes artifacts that purely structural constraints cannot prevent, particularly in ribs and abdominal soft tissue.","The tissue-aware and respiration-aware registration losses provide a template for handling whole-body deformation, including the discontinuous rib-liver sliding that uniform smoothness regularization would over-smooth."],"supporting_citations":[{"why":"Defines the unpaired cycle-consistency translation approach that the proposed framework must beat and that motivates the need for explicit registration.","marker":"[18]"},{"why":"Supplies the structure-constrained baseline used for head-to-head PSNR, SSIM, and PET SUV comparisons.","marker":"[16]"},{"why":"Supplies the gradient-consistency baseline used in the whole-body synthesis comparison.","marker":"[17]"},{"why":"Provides the 3D U-Net architecture on which the synthesis generator and registration network are built.","marker":"[41]"},{"why":"Supplies the registration-network design and the similarity-plus-smoothness loss template that SpatAlign modifies.","marker":"[44]"},{"why":"Provides the differentiable mutual-information neural estimator used in the tissue-aware similarity loss for registration.","marker":"[47]"},{"why":"Supplies the pre-trained multi-organ segmenter that defines tissue weights and contrastive organ pairs for semantic alignment.","marker":"[49]"},{"why":"Provides the contrastive loss formulation used to pull same-organ features together and push different organs apart.","marker":"[51]"},{"why":"Defines the conventional UTE segmentation-based attenuation-correction method compared in the PET SUV evaluation.","marker":"[11]"}],"fun_headline_variants":["Whole-body MR-to-CT gets a dual alignment boost","Dual alignment is key to whole-body MR-to-CT synthesis","Spatial and semantic alignment sharpens PET/MR synthesis","Organ-aware contrastive learning improves synthetic CT","Structure-guided synthesis aligns organ semantics for PET"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole pipeline assumes the pre-trained whole-body organ segmenter produces reliable organ masks on both real CT and synthetic CT; if those masks are wrong, the tissue weights in registration and the positive and negative pairs in contrastive learning are noisy, and the paper itself notes that occasional poor segmentation can compromise contrastive learning.","fun_headline_variants_meta":{"raw":{"variants":["Whole-body MR-to-CT gets a dual alignment boost","Dual alignment is key to whole-body MR-to-CT synthesis","Spatial and semantic alignment sharpens PET/MR synthesis","Organ-aware contrastive learning improves synthetic CT","Structure-guided synthesis aligns organ semantics for PET"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001143,"raw_usage":{"total_tokens":4753,"prompt_tokens":964,"completion_tokens":3789,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":580,"completion_tokens_details":{"reasoning_tokens":3713}},"tokens_in":580,"tokens_out":3789,"duration_ms":46336,"temperature":1.0,"reasoning_tokens":3713,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:04:08.229600+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same region-wise PSNR, SSIM, and PET SUV-difference evaluation on an independent multi-center whole-body PET/MR dataset with paired CT: if the framework's whole-body PSNR advantage over the strongest unpaired-translation baseline disappears, or if its spine and femur SUV errors are not lower than the segmentation-based UTE method, the claimed superiority is refuted. A more targeted check is to corrupt or remove the organ-segmentation masks during training; if the framework's gain over the no-semantic-alignment ablation vanishes, the semantic-alignment claim is the operative cause.","supporting_citations":[{"cited_title":"Unsupervised mr-to-ct synthesis using structure-constrained cyclegan,","cited_arxiv_id":null,"evidence_quote":"Supplies the structure-constrained baseline used for head-to-head PSNR, SSIM, and PET SUV comparisons."},{"cited_title":"Cross-modality image synthesis from unpaired data using cyclegan: Effects of gradient consistency loss and training data size,","cited_arxiv_id":null,"evidence_quote":"Supplies the gradient-consistency baseline used in the whole-body synthesis comparison."},{"cited_title":"Mutual information neural estimation for unsupervised multi-modal registration of brain images,","cited_arxiv_id":null,"evidence_quote":"Supplies the pre-trained multi-organ segmenter that defines tissue weights and contrastive organ pairs for semantic alignment."},{"cited_title":"Mri-based attenuation correction for pet/mri using ultrashort echo time sequences,","cited_arxiv_id":null,"evidence_quote":"Defines the conventional UTE segmentation-based attenuation-correction method compared in the PET SUV evaluation."}],"review_version":1}