{"id":"a901348a-2ee9-4217-873f-23fd779d9fcb","arxiv_id":"2507.00206","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Med-LSDM generates 3D medical volumes from semantic maps in a VQ-GAN latent space and reports Dice close to real data on breast MRI.","lead":"This paper builds a 3D medical image generator, Med-LSDM, that creates MRI or CT volumes from segmentation maps using a latent diffusion model in a compressed VQ-GAN space. It reports small differences between synthetic and real images on breast MRI, which could support privacy-preserving data sharing and augmentation if validated.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported 3D-FID of 0.0054 is almost certainly computed with 10 real and 10 synthetic volumes; with high-dimensional Med3D features the FID estimate is severely downward-biased, so the paper's central small-domain-gap evidence is not yet established.","rationale":"The paper's central claim is that Med-LSDM produces 3D volumes with a small domain gap to real data and is useful for data augmentation. The reader correctly notes the VQ-GAN latent fidelity is assumed rather than quantified. I agree that is a real assumption, but it is downstream of a more immediate problem: the quantitative evidence used to establish the domain gap is not trustworthy as reported. The Duke Breast test set has 10 labeled scans; the FID section gives no sample count or feature dimension, so the reported 3D-FID of 0.0054 is consistent with a small-sample, high-dimensional FID artifact rather than a true distributional match. If this number is unreliable, the 'small domain gap' conclusion is unsupported, and the VQ-GAN fidelity concern becomes moot because the metrics that would validate it are invalid. The Dice result, while suggestive of mask faithfulness, is not a test of augmentation utility; the abstract's statement that synthetic data 'are useful for data augmentation' is never directly tested. The paper also provides no code or data, so the numbers cannot be checked. I therefore recommend the verdict move from CONDITIONAL to UNVERDICTED: the central empirical claims cannot be evaluated from the manuscript as written. The condition for a future positive verdict is not merely latent-fidelity reporting; it is a recomputed 3D-FID with adequate sample size and an actual data-augmentation experiment.","tokens_in":14259,"tokens_out":8282,"duration_ms":92964,"concrete_test":"Recompute the Duke Breast 3D-FID using at least 100 real and 100 synthetic volumes (sampling multiple latents per test semantic map), explicitly reporting the number of samples n and the Med3D feature dimension d, with bootstrap confidence intervals and a shuffled-label null. If the recomputed value is substantially above 0.0054 or its confidence interval overlaps the null, the small-domain-gap conclusion is unsupported.","verdict_should_be":"UNVERDICTED","load_bearing_attack":"The central quantitative evidence for the paper's headline 'small domain gap' is the Duke Breast 3D-FID of 0.0054 (Table 3). The experimental setup states that only 10 labeled Duke Breast scans form the test set, and synthetic images are generated from the semantic maps of that test set (Section 4.2); the paper does not state how many volumes were used for the FID calculation or the dimensionality of the Med3D features. FID is a plug-in estimate of the Fréchet distance between two Gaussians fitted to features; with n=10 volumes and features of dimension d much larger than n, the estimate is strongly biased toward zero and has enormous variance. The contrast between 2D FID (193.620) and 3D-FID (0.0054) on the same data is consistent with this artifact. The Dice comparison (0.70964 vs 0.71496) is not a substitute: it measures agreement of a real-trained segmentation network with the conditioning mask on synthetic images, not distributional similarity, and it does not test the abstract's claim that synthetic data 'are useful for data augmentation,' which is never evaluated with an actual augmentation experiment. If the 3D-FID estimate is unreliable, the 'small domain gap' conclusion collapses; if the augmentation claim is untested, the practical utility claim is unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Med-LSDM, a 3D latent semantic diffusion model for medical image synthesis. A 3D VQ-GAN is first trained to compress volumetric images into a discrete latent space; a semantic diffusion model is then trained in that latent space, using SPADE modules conditioned on semantic maps to guide denoising. The authors evaluate the model on AutoPET, SynthRAD2023, and Duke Breast datasets, comparing against a 2D segmentation-guided diffusion baseline (SegGuidedDiff). The central claims are that Med-LSDM achieves a very small distributional gap with real data (3D-FID 0.0054 on Duke Breast) and that the synthetic data are useful for data augmentation, based on a Dice similarity of 0.70964 between a real-trained segmentation network's output on synthetic images and the input masks, versus 0.71496 on real test images.","tokens_in":14513,"tokens_out":3408,"duration_ms":39648,"significance":"If the claims are substantiated, this would be a useful contribution to 3D medical image synthesis, offering a way to generate volumetric images from semantic maps in latent space, which could reduce computational cost and support privacy-preserving data sharing and augmentation. The architecture is clearly described, the two-stage training (VQ-GAN followed by latent diffusion) is sound in principle, and the paper includes a baseline comparison and qualitative volumetric visualizations. However, the central quantitative evidence for the 'small domain gap' and 'useful for augmentation' claims is not yet convincing: the 3D-FID estimate is prone to severe bias at the reported sample size, no actual augmentation experiment is performed, and the VQ-GAN reconstruction fidelity is only qualitatively assessed. These issues need to be addressed before the main conclusions can be accepted.","major_comments":[{"comment":"The 3D-FID results, including the headline value of 0.0054 on Duke Breast, are not reliable as reported. The experimental setup states that the Duke Breast test set contains only 10 scans, and the synthetic images are generated from the semantic maps of that same test set. If the FID is computed on these 10 real and 10 synthetic volumes using high-dimensional Med3D features, the plug-in Fréchet distance estimate is strongly biased toward zero. The paper does not report the number of volumes used for the FID computation nor the feature dimensionality. The stark contrast between the 2D FID (193.620) and 3D-FID (0.0054) on the same data is consistent with this artifact. Please report the exact sample size, feature dimension, and ideally a bootstrap confidence interval or an unbiased estimator. Without this, the 'small domain gap' conclusion (abstract, Section 5) is not established.","section":"Section 4.1 and Section 4.4, Table 3"},{"comment":"The abstract and Section 5 claim that the synthetic data are 'useful for data augmentation,' but no augmentation experiment is conducted. Table 2 reports Dice scores of a segmentation network trained on real Duke Breast data when evaluated on synthetic images. This is a mask-faithfulness check, not a test of whether adding synthetic images to a training set improves downstream task performance. To support the augmentation claim, the authors should train segmentation networks on (i) real data only and (ii) real plus synthetic data, and compare performance on a held-out real test set. This is a standard and necessary experiment for the stated practical utility.","section":"Section 4.2 and Abstract"},{"comment":"The assumption that the VQ-GAN latent space 'can effectively compress spatial information while preserving its correspondence with the pixel space' is load-bearing for the entire generation chain, but it is supported only by visual inspection (Figure 7). No quantitative reconstruction error is reported (e.g., PSNR/SSIM on held-out volumes, or segmentation Dice on reconstructed images). Given that a 4x compression is used, the authors should provide quantitative evidence that clinically relevant structures survive the compression step. Without this, the possibility that the diffusion model operates on an overly smoothed or anatomically altered latent space cannot be ruled out.","section":"Section 4.3"}],"minor_comments":[{"comment":"There are numerous typos and grammatical errors, e.g., 'Trasnlation' (Section 2 heading), 'Econder' (Section 3.3), 'comparation' (Figure 6), 'advant a' (Section 4.4), 'Med-LSTM' (should be 'Med-LSDM', Section 2), and inconsistent capitalization. A careful proofread is needed.","section":"Throughout"},{"comment":"The Duke Breast row lists '3D Scans (Train/Test)' as '(822 + 90)/10'. The text explains that 822 unlabeled scans are used for VQ-GAN training and 90 labeled scans for the diffusion model, but this notation is confusing. Please separate the two training subsets explicitly.","section":"Table 1"},{"comment":"The abstract reports a Dice score of 0.71496 for real test data, while Table 2 reports 0.715. Please make the values consistent, and report them to the same number of significant digits in both places.","section":"Table 2 and Abstract"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a relevant problem and the general pipeline is plausible, but the main claims currently rest on metrics that are either unreliable at the reported sample size or not measuring what is claimed. The revisions required are substantial but feasible: a proper FID evaluation, an actual augmentation study, and quantitative reconstruction assessment. I would support reconsideration after these are provided."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe genuinely new thing here is the architecture: the authors are the first to run SPADE-conditioned diffusion inside a 3D VQ-GAN latent space for medical volumes. The combination is clean and the 3D setup is sensible. The paper also does something right that many medical synthesis papers skip: it compares against a 2D slice-wise baseline (SegGuidedDiff) and shows better inter-slice consistency in coronal and sagittal views. The authors are honest about the SynthRAD cross-modal failure and about the 32-slice limit.\n\nThat said, the quantitative support for the headline claim is thin. The 3D-FID of 0.0054 on Duke Breast is almost certainly computed from 10 real and 10 synthetic volumes. With high-dimensional Med3D features, that plug-in estimate is severely biased toward zero. The contrast between 2D FID (193.6) and 3D-FID (0.0054) on the same data is consistent with this artifact, not with a miraculously small domain gap. The Dice comparison (0.710 vs 0.715) is a re-segmentation consistency check with a real-trained network; it shows the generator respects masks, but it does not test whether synthetic data help augmentation. No actual augmentation experiment exists, despite the abstract's claim that the results demonstrate usefulness for data augmentation.\n\nThe VQ-GAN latent fidelity is also only defended by a few visual examples; there is no quantitative reconstruction error, so the whole conditioned-diffusion chain rests on an unmeasured assumption. Minor issues: the paper has typos (\"Trasnlation\", \"Med-LSTM\") and the baseline Med-DDPM is not compared numerically.\n\nIf the authors add a real augmentation experiment (train on real+synthetic vs real-only), compare against Med-DDPM, report confidence intervals, and release code/data, the claims would be substantially stronger. As is, the central quantitative evidence is not established.\n\nWho this is for: people working on medical image synthesis and privacy-preserving data sharing. It deserves a serious referee because the direction matters and the combination is new, but it needs major revision before acceptance. I would send it out, not desk-reject it.","headline":"Med-LSDM is a new 3D latent semantic synthesis pipeline, but its headline numbers rest on an unreliable FID estimate and the augmentation claim is untested.","tokens_in":15074,"tokens_out":1911,"would_cite":false,"duration_ms":21851,"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":"This paper claims that 3D medical volumes generated from semantic maps by a latent diffusion model are close enough to real scans to serve as privacy-preserving training data.","keywords":["3D medical image synthesis","latent diffusion model","semantic image synthesis","VQ-GAN","data augmentation","privacy preservation","denoising diffusion probabilistic models","medical imaging"],"falsifier":"Run the same segmentation network on real test volumes that have been passed through the VQ-GAN encoder and decoder without any diffusion; if the Dice score on these reconstructed volumes drops well below 0.715, then the 4x compression, not the diffusion model, is what limits anatomical fidelity, and the small-domain-gap result would not transfer to tasks needing finer detail.","tokens_in":14032,"feed_emoji":"🩻","tokens_out":5324,"duration_ms":54185,"temperature":0.7,"pith_summary":"This paper proposes Med-LSDM, a latent diffusion model that synthesizes full 3D medical volumes from semantic segmentation maps instead of generating 2D slices and stacking them. The authors aim to show that synthetic volumes generated this way lie close enough to real patient data to serve as privacy-preserving data augmentation. Their evidence is a 3D-FID of 0.0054 on the conditional Duke Breast MRI dataset and a downstream segmentation Dice score of 0.70964 on synthetic volumes versus 0.71496 on real test volumes. The practical promise is that institutions could share de-identified semantic maps and distribute synthetic images rather than raw patient scans, easing data scarcity and privacy restrictions.","feed_headline":"3D diffusion model makes synthetic MRI nearly match real scans","feed_subtitle":"Semantic-map-conditioned latent generation yields near-real Dice for scarce medical data.","key_machinery":"Med-LSDM couples a pre-trained 3D VQ-GAN, which compresses a 256x256x32 volume into a discrete latent code of size 64x64x8 with a 16384-entry codebook, with a semantic diffusion model that denoises Gaussian noise in that latent space conditioned on the semantic map. The conditioning is done by semantic diffusion decoder residual blocks that use SPADE-style spatially-adaptive normalization (modulating normalized feature maps with learned scale and shift from the semantic map). After denoising, the latent code is quantized against the frozen codebook and decoded to the voxel space, which is the step that carries the claim that 4x compression preserves clinically relevant structure.","core_discovery":"The paper's central claim is that running a denoising diffusion model inside the learned latent space of a 3D VQ-GAN, with the semantic map injected at multiple decoder stages via spatially-adaptive normalization, produces volumetric medical images whose distribution and anatomy match real data closely enough for downstream use. The authors report that segmentation performance on Med-LSDM-generated volumes (Dice 0.70964) is nearly identical to segmentation performance on real test volumes (Dice 0.71496), and far above the 2D slice-by-slice baseline (0.602). They position this as the first 3D semantic medical image synthesis in latent space, with the main remaining weakness being behavior in limited-data cross-modal settings like SynthRAD2023.","pith_inferences":["A direct test the paper leaves implicit: measure VQ-GAN reconstruction fidelity quantitatively (e.g., Dice between original and reconstructed real volumes); if that reconstruction Dice is much lower than 0.71, the latent space itself, not the diffusion, may be the bottleneck.","The reported 3D-FID of 0.0054 is unusually low compared with typical FID scales, suggesting the Med3D feature distribution is narrow; a reader should compare models on the same feature extractor before treating absolute numbers as meaningful.","The method can be extended to unconditional or weakly supervised settings by replacing the semantic map with a predicted or partial label, which would broaden the privacy-preserving claim beyond datasets that already have annotations.","Another testable extension is an augmentation experiment where a segmentation model is trained on real plus synthetic volumes and compared against real-only training; the paper demonstrates equivalence of synthetic to real but does not directly demonstrate a gain from adding synthetic data."],"forward_implications":["If correct, researchers can train segmentation networks on synthetic volumes and expect test-time behavior close to real-data training, making augmentation viable when real labels are scarce.","De-identified semantic maps become a shareable asset: only maps and synthetic volumes need leave an institution, reducing re-identification risk.","3D consistency across slices is preserved, so volumetric analyses (coronal and sagittal views) avoid the inter-slice artifacts of stacked 2D generation.","On cross-modal tasks with small datasets, the method's advantage shrinks, so future work needs modality-aware conditioning or pretraining to close that gap."],"supporting_citations":[{"why":"Provides the VQ-GAN architecture whose latent space the diffusion model operates in.","marker":"[14]"},{"why":"Defines the denoising diffusion probabilistic model and the simplified noise-prediction objective used for training.","marker":"[7]"},{"why":"Supplies the SPADE spatially-adaptive normalization mechanism used to condition denoising on semantic maps.","marker":"[17]"},{"why":"Is the 2D slice-wise SegGuidedDiff baseline Med-LSDM is compared against and outperforms on 3D consistency and Dice.","marker":"[26]"},{"why":"Is the prior conditional 3D diffusion approach whose conditioning at the initial noise stage the paper improves on.","marker":"[28]"},{"why":"Establishes latent diffusion as the efficiency approach Med-LSDM adapts to 3D medical volumes.","marker":"[22]"},{"why":"Defines the 3D-FID metric used to measure distributional gap between real and synthetic volumes.","marker":"[35]"},{"why":"Provides the pre-trained Med3D network whose features are used to compute 3D-FID.","marker":"[36]"},{"why":"Duke Breast dataset, the main benchmark where the headline Dice and 3D-FID numbers are reported.","marker":"[33]"}],"fun_headline_variants":["3D latent diffusion yields synthetic medical images near real","Latent 3D diffusion matches real scans in segmentation test","Med-LSDM: 3D synthetic images close the gap with real data","Synthetic 3D medical images rival real ones for segmentation","3D semantic diffusion model generates near-real medical volumes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole approach assumes that squeezing each volume down to a quarter of its original size keeps all clinically meaningful details; the paper shows only pictures to support this, so if fine structures get lost in that squeeze, the generated images will look smooth but cannot be trusted for clinical tasks.","fun_headline_variants_meta":{"raw":{"variants":["3D latent diffusion yields synthetic medical images near real","Latent 3D diffusion matches real scans in segmentation test","Med-LSDM: 3D synthetic images close the gap with real data","Synthetic 3D medical images rival real ones for segmentation","3D semantic diffusion model generates near-real medical volumes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000186,"raw_usage":{"total_tokens":1313,"prompt_tokens":923,"completion_tokens":390,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":539,"completion_tokens_details":{"reasoning_tokens":304}},"tokens_in":539,"tokens_out":390,"duration_ms":4791,"temperature":1.0,"reasoning_tokens":304,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:21:49.870098+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same segmentation network on real test volumes that have been passed through the VQ-GAN encoder and decoder without any diffusion; if the Dice score on these reconstructed volumes drops well below 0.715, then the 4x compression, not the diffusion model, is what limits anatomical fidelity, and the small-domain-gap result would not transfer to tasks needing finer detail.","supporting_citations":[{"cited_title":"Mazurowski","cited_arxiv_id":null,"evidence_quote":"Is the 2D slice-wise SegGuidedDiff baseline Med-LSDM is compared against and outperforms on 3D consistency and Dice."},{"cited_title":"Conditional diffusion models for semantic 3d brain mri synthesis","cited_arxiv_id":null,"evidence_quote":"Is the prior conditional 3D diffusion approach whose conditioning at the initial noise stage the paper improves on."},{"cited_title":"Hierarchical amortized gan for 3d high resolution medical image synthesis","cited_arxiv_id":null,"evidence_quote":"Defines the 3D-FID metric used to measure distributional gap between real and synthetic volumes."},{"cited_title":"Med3d: Transfer learning for 3d medical image analysis, 2019","cited_arxiv_id":null,"evidence_quote":"Provides the pre-trained Med3D network whose features are used to compute 3D-FID."},{"cited_title":"Saha, M.R","cited_arxiv_id":null,"evidence_quote":"Duke Breast dataset, the main benchmark where the headline Dice and 3D-FID numbers are reported."}],"review_version":1}