{"id":"7e07eae7-ed9e-4822-a2d2-24ef9e37b3ee","arxiv_id":"2412.16050","paper_version":4,"verdict":"REJECT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"SF-VD generates labeled fluoroscopy videos via separate scene and motion diffusion models and reports improved guidewire segmentation, but test-set contamination weakens the evidence.","lead":"A two-stage diffusion model generates synthetic fluoroscopy videos with guidewire masks and uses them to augment segmentation training data. The paper reports improved guidewire segmentation, but the evaluation is undermined by a train/test overlap: the generator was trained on the same videos later used as the test set.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Generative models are trained on the entire F-video dataset, including the segmentation test split; the reported Dice gains may reflect test-set leakage rather than augmentation quality.","rationale":"The reader's REJECT is justified, but the weakest assumption they list (cardiac motion being simple enough for a 2D conditional diffusion model) is not the most load-bearing issue. The decisive problem is that the generative models are trained on the full F-video corpus and the segmentation test set is a split of the same corpus, so the synthetic augmentation pipeline is transductive with respect to the test videos. With a small dataset of 400 videos, a diffusion model can memorize or closely approximate training frames, and the 1200 synthetic videos could carry test-video pixel statistics into the segmentation training set. This would inflate Dice and HD in Table 2 regardless of whether the motion-distribution model is adequate. The paper's FVD/DS/OS metrics do not address this because they compare generated videos with the same corpus used for training. Additional reproducibility problems (no released code/data, and Appendix E pseudocode that swaps the conditional and unconditional noise terms relative to Eqs. 4, 6, and 8) make it impossible to reconstruct the exact method. I credit the paper for including multiple segmentation backbones, specialized guidewire models, ablations of frame-consistency and segmentation-guidance, and comparisons against other augmentation strategies; those are appropriate and useful. However, the central empirical claim is not cleanly supported as presented. The proposed retraining check would settle whether the improvement survives without test-set exposure; if it does, the claim would become credible. Until then, the reader's REJECT should stand, so the verdict is UNCHANGED.","tokens_in":19121,"tokens_out":4981,"duration_ms":46018,"concrete_test":"Retrain θs and θm using only the 80% segmentation-training F-video split (plus P-image for θs), regenerate the 1200 synthetic videos from masks drawn only from that split, and rerun the Table 2 comparison on the untouched 10% test split. If Dice gains over 'No aug' remain within ~0.02 of 0.61 and HD stays within ~1 px of the reported values, leakage is not the driver; if gains drop materially (e.g., Dice < 0.56 or HD > 6 px), the published improvement is inflated by test-set exposure. As a complementary check, compute nearest-neighbor pixel similarity between synthetic training videos and test frames to quantify memorization.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Central claim: SF-VD synthetic augmentation raises guidewire-segmentation Dice from ~0.52 to ~0.61 (Table 2). A necessary condition is that the gain comes from the generator's ability to synthesize useful new data, not from its access to test-set statistics. The paper trains the scene model θs on both P-image and F-video and the motion model θm on F-video, then splits the same F-video corpus 80/10/10 for segmentation training/validation/test (Experiments). The diffusion models therefore see every pixel of the held-out test videos. With only 400 16-frame videos at 512x512, a conditional diffusion model can plausibly memorize or closely mimic frames, so 1200 synthetic training videos can leak test backgrounds and appearances into the segmentation training set. The reported Dice/HD gains are then not clean evidence for label-efficient augmentation. No code or data is released, and Appendix E Algorithm 3 contradicts Eqs. 4, 6, and 8 by swapping unconditional/conditional noise terms and using opposite guidance weights, so the method as published is not unambiguously reproducible. The reader's 'motion is simple' assumption is secondary; the decisive problem is test-set exposure.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SF-VD, a data-augmentation framework that trains two 2D diffusion models to synthesize labeled cardiac fluoroscopy videos for guidewire segmentation: a scene model generates the leading frame conditioned on a guidewire mask, a motion model generates subsequent frames, and a frame-consistency sampling stage generates intermediate frames between already-sampled endpoints. A segmentation network guides the reverse process to vary wire contrast. The authors train and evaluate downstream generic and specialized segmentation models with and without augmentation, reporting Dice/HD gains together with FVD/DS/OS video-quality comparisons and ablations of the two proposed components.","tokens_in":19328,"tokens_out":12136,"duration_ms":111829,"significance":"The paper addresses a practical problem, annotation-efficient training for thin guidewire segmentation in fluoroscopy, and the two-model decomposition of scene and motion distributions is a reasonable inductive bias for sharing unlabeled image data with a small fully annotated video corpus. The paper also evaluates across many architectures and includes ablations. However, the central empirical claim is not currently supported: the generative models and the segmentation-guiding network are trained on the same F-video corpus from which the segmentation validation/test splits are later drawn, so the reported Dice improvements (e.g., 0.52 to 0.61 for Dynvnet in Table 2) may reflect leakage rather than augmentation quality. If re-run with a clean protocol, the methodology could be valuable, but as submitted the quantitative evaluation is not trustworthy.","major_comments":[{"comment":"The evaluation protocol gives the generative models access to the segmentation test and validation videos. The scene model theta_s is trained on the P-image and F-video datasets, and the motion model theta_m is trained on the F-video dataset, while the downstream segmentation evaluation splits the same F-video corpus into 80/10/10 (Experiments, 'Wire Segmentation Evaluation'; Appendix B.3). Moreover, the segmentation network psi used for guidance is trained on 'all annotated frames from the P-image and F-video dataset', which includes the held-out frames. Hence the 1,200 synthetic training videos can contain the exact backgrounds, anatomy, and wire contrast of the validation/test videos, and with 400 16-frame videos at 512 by 512, near-copy or memorized frames are plausible. The Dice/HD improvements in Table 2 are therefore not clean evidence that synthetic augmentation transfers to unseen data. The experiments need to be rerun with theta_s, theta_m, and psi trained only on the segmentation training split, the validation split used only for model selection, and the test split never involved in any generative training.","section":"Experiments, 'Wire Segmentation Evaluation'; Appendix B.3"},{"comment":"Algorithm 3 is inconsistent with the equations and with the text's guidance weights. Stage 1 line 8 gives epsilon_bar = (1 - omega) epsilon_mask + omega epsilon_no-mask, which is the opposite assignment to Eq. (4); Stage 2 line 18 does the same relative to Eq. (6); and Stage 3 line 28 defines epsilon_no-frame as theta_m(XS, s, M, F1), although Eq. (8) requires the mask-only function theta_m(XS, s, M), so the F1 condition is double-counted in line 30. The text states omega = -2.5 for Eq. (6) and -1.5 for Eq. (8), but Algorithm 3 does not set these values, and the swapped signs make the actual weighting ambiguous. Finally, line 29 computes a variance-like quantity multiplied by a gradient of M to X rather than applying the mean modification gamma * sigma^2 * grad log p_psi(M|x) from Eq. (9). Because no code is released, these inconsistencies make the published method not unambiguously reproducible.","section":"Appendix E, Algorithm 3; Eqs. (4), (6), (8), and (9)"},{"comment":"The text claims that the simultaneous application of FC and SG yields the best results for Dice, but Table 3 reports Dice 0.60 for FC-only versus 0.59 for FC+SG, so the best Dice is achieved without SG. SG does improve HD, G2RE, and R2GE, but the summary sentence should be corrected to say that the combination yields the best boundary/distance metrics rather than the best Dice.","section":"Appendix D, Table 3; Ablation Study"}],"minor_comments":[{"comment":"The text says SF-VD achieves the 'highest FVD,' but FVD is labeled with a down arrow and SF-VD has the lowest value in Table 1; please rephrase.","section":"Results, paragraph after Table 1"},{"comment":"Please specify the total number of gradient updates or iterations for each training setting, since reducing epochs from 1000 to 250 when the augmented dataset is more than four times larger is only a fair comparison if the total optimization budget is matched.","section":"Appendix B.3"},{"comment":"The comment 'Stage 1: Leading Frame Synthesis' inside Algorithm 2 appears to be a copy-paste artifact; Algorithm 2 trains the segmentation network.","section":"Appendix E, Algorithm 2"},{"comment":"The symbol F is used for a single leading frame in Eqs. (5)-(6) and for leading/concluding frames F1/F2 in Eqs. (7)-(8); please unify the notation.","section":"Methodology, Eqs. (5)-(8)"},{"comment":"The paper does not include a data/code availability statement; the private dataset and unreleased code further limit reproducibility.","section":"General"}],"recommendation":"reject","confidential_remarks":"I want to flag for the editor that the protocol overlap is the decisive issue: every quantitative segmentation result is produced with models that saw the test videos, and the segmentation-guidance network was trained on test annotations. No code or data is provided, and the appendix pseudocode contradicts the equations, so the reported numbers cannot be independently checked. The underlying idea may be salvageable, but the current submission does not provide valid evidence for its central claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: SF-VD is a method-level contribution with a real problem in its evaluation. The authors decompose fluoroscopy video generation into scene and motion using two 2D diffusion models, add frame-consistency sampling, and use a segmentation network to guide wire contrast. That combination is new for fluoroscopy, and the ablations indicate each piece contributes. The reported segmentation improvements are consistent across eight architectures.\n\nThe soft spot is the data protocol. The generative models are trained on the full F-video dataset, and the same corpus is later split 80/10/10 for the segmentation experiment. The generator therefore has already seen the test videos. With only 400 16-frame clips at 512x512, a diffusion model can memorize or closely imitate test frames, so synthetic augmentation can carry test-set backgrounds and patient appearance into the training set. That directly undermines the Dice and HD gains in Table 2. The right fix is to train the generator only on the 80% training split, or to hold out whole patients from both generator and segmenter. No code or data is released to re-check.\n\nThere is also a reproducibility wrinkle: Appendix E Algorithm 3 swaps the unconditional/conditional noise terms relative to Eqs. 4, 6, and 8, and uses guidance weights that are not fully defined. A reimplementation would get stuck.\n\nThat said, the method itself is plausible. Separate scene and motion modeling fits cardiac fluoroscopy, where the anatomy is roughly static and wires move. The motion-simplicity assumption is defensible. The segmentation-guided contrast mechanism is a legitimate conditioning choice, not a circular argument. The paper also honestly notes limitations like slow generation and small training data.\n\nWho should read this: researchers working on medical video synthesis or guidewire segmentation. The idea is worth discussing, but the headline numbers should not be trusted. If this crossed my desk I would send it to peer review because the method merits scrutiny, but I would expect a major revision focused on the evaluation protocol, and I would not accept the current empirical claims.\n\nRecommendation: engage with the method, not with Table 2 as evidence.","headline":"SF-VD's two-stage diffusion pipeline for fluoroscopy video synthesis is clever, but training the generator on the test split makes the reported segmentation gains untrustworthy.","tokens_in":19860,"tokens_out":4893,"would_cite":false,"duration_ms":42201,"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":"By splitting video generation into scene and motion, a new diffusion method synthesizes labeled fluoroscopy videos that lift guidewire segmentation Dice to around 0.61.","keywords":["guidewire segmentation","cardiac fluoroscopy","video diffusion model","data augmentation","label-efficient learning","classifier-free guidance","frame-consistency sampling","medical video synthesis"],"falsifier":"Train SF-VD on fluoroscopy videos that violate the simple-motion assumption, such as rapid table panning, large out-of-view device motion, or significant breathing-induced organ displacement, then measure the downstream Dice of a fixed segmentation network trained with the generated videos; if the augmentation no longer improves over no augmentation, the motion-model assumption is the failure point. A second, more direct falsifier is to check whether the frame-consistency sampling actually prevents temporal artifacts by computing FVD on videos generated from held-out mask sequences and asking human raters to detect frame-to-frame anatomical discontinuities.","tokens_in":18924,"feed_emoji":"🩻","tokens_out":9466,"duration_ms":73807,"temperature":0.7,"pith_summary":"The paper aims to show that the shortage of annotated guidewire videos in cardiac fluoroscopy can be sidestepped by generating synthetic training videos with a diffusion model that is label-efficient. Its central move is to split the hard joint distribution of a video given wire masks into a scene distribution for the first frame and a motion distribution for subsequent frames, each modeled by a separate 2D diffusion network. This split lets the scene model train on a mixture of fully annotated videos and thousands of unannotated frames, so the total annotation requirement stays at about 400 fully labeled videos. The paper's main evidence is that augmenting segmentation training with these synthetic videos raises Dice from about 0.52 to about 0.61 for generic networks such as Dynvnet, SwinUnetr, and MLP-Mixer, and improves or matches wire-specific models like PARN, RTGN, and Ariadne+. If the claim holds, it means useful data augmentation for thin-structure medical segmentation can be produced from very limited labeled video.","feed_headline":"Synthetic videos push guidewire segmentation Dice from 0.52 to 0.61","feed_subtitle":"SF-VD splits video generation into scene and motion, needing only 400 real annotated videos for training.","key_machinery":"The load-bearing object is the distribution decomposition $P(V \\mid M) \\approx P(I_0 \\mid M_0)\\,P(I_1,\\ldots,I_{N-1} \\mid M_1,\\ldots,M_{N-1}, I_0)$, realized as two 2D conditional DDPMs with classifier-free guidance. The scene model $\\theta_s$ synthesizes the leading frame from a wire mask; the motion model $\\theta_m$ synthesizes subsequent frames from a mask plus one or two reference frames, and is used both for the concluding frame and, with frame-consistency sampling, for intermediate frames. The segmentation-guided mechanism injects the gradient of a wire-segmentation network into the reverse process, scaling wire contrast by a strength $\\gamma$ chosen per sample.","core_discovery":"The central discovery is that the conditional distribution of a fluoroscopy video given a sequence of guidewire masks can be split productively into two 2D conditional diffusion problems rather than one 3D video diffusion problem. A scene-distribution model generates the leading frame from a mask, and a motion-distribution model generates every later frame from the leading frame (or from both endpoint frames) plus the corresponding masks. The motion model also takes the frame distance as input, and a frame-consistency sampling procedure generates intermediate frames conditioned on both a leading and a concluding frame, which the ablation shows is what suppresses jittery background motion. A separate segmentation network guides the reverse diffusion to modulate guidewire contrast, increasing appearance diversity. On a private hospital dataset, the authors report that the resulting synthetic videos improve Dice from about 0.52 to about 0.61 for three generic segmentation networks and improve or match wire-specific segmentation models, with the best results coming from combining frame-consistency and segmentation-guided mechanisms.","pith_inferences":["Inference, beyond the paper: the scene-plus-motion split appears generic to thin interventional devices, but the paper only demonstrates it for guidewires; a natural extension is to catheters, stents, and ablation tips, where annotated video is similarly scarce.","Inference: because the evaluation uses a private, single-center dataset, the reported Dice gains may not transfer across X-ray vendors or institutions; a multi-center study would be needed to establish generalizability.","Inference: the 40-second per-video generation cost on an A100, stated as a limitation in the paper, makes large-scale augmentation expensive; distilling each 2D diffusion model into a consistency model would likely reduce this by an order of magnitude.","Inference: the paper does not measure how segmentation performance varies with the number of synthetic videos; we expect a saturation point exists, and identifying it would let practitioners match generation budget to annotation budget."],"forward_implications":["Augmenting with SF-VD synthetic videos raises Dice of generic 3D segmentation networks (Dynvnet, SwinUnetr, MLP-Mixer) from about 0.52 to about 0.61, and reduces Hausdorff distance by a factor of roughly three to seven compared with unaugmented training.","SF-VD outperforms or matches both a 3D video diffusion baseline (VD) and a latent video diffusion baseline (LVDM) on video realism (FVD), diversity, and downstream segmentation, while producing more novel content as measured by Overfitting Score.","Frame-consistency sampling is what improves region-overlap metrics, while segmentation-guided contrast modulation is what reduces distance-based errors; combining both yields the best overall segmentation, so each mechanism contributes a distinct benefit.","Because the scene model can be trained with unannotated images, the effective annotation cost of the method is concentrated in the motion model, which needs only fully labeled videos—suggesting that unannotated fluoroscopy archives can be leveraged for augmentation.","The paper positions SF-VD as the first fluoroscopy video synthesis approach for guidewire segmentation augmentation, so if its results replicate, it provides a new tool for label-efficient training in interventional imaging."],"supporting_citations":[{"why":"Supplies the DDPM forward/reverse process that both the scene and motion models are built on.","marker":"(Ho, Jain, and Abbeel 2020)"},{"why":"Provides the predicted variance and the training objective (IDDPM) used in Eq. 1 and Appendix A.","marker":"(Nichol and Dhariwal 2021)"},{"why":"Gives the classifier-free guidance formulation used to condition the scene model on masks and the motion model on frames.","marker":"(Ho 2022)"},{"why":"Supplies the classifier-guidance style used by the segmentation-guided mechanism to modulate wire contrast (Eq. 9).","marker":"(Dhariwal and Nichol 2021)"},{"why":"Defines the Video Diffusion Model baseline (VD) that SF-VD is compared against and outperforms on FVD and segmentation.","marker":"(Ho et al. 2022)"},{"why":"Defines the Latent Video Diffusion Model baseline (LVDM) that SF-VD is compared against and beats on diversity and downstream Dice.","marker":"(He et al. 2022)"},{"why":"Provides the nnU-Net/Dynvnet architecture used both as a downstream segmentation network and as the backbone for the segmentation-guided mechanism.","marker":"(Isensee et al. 2020)"},{"why":"Supplies PARN, a wire-specific segmentation network whose performance with SF-VD augmentation is the paper's evidence for benefit on tailored architectures.","marker":"(Zhou et al. 2020)"},{"why":"Supplies Ariadne+, another wire-specific baseline whose improvement with SF-VD augmentation is reported.","marker":"(Caporali et al. 2022)"},{"why":"Defines FVD, the video quality metric used to compare SF-VD against VD and LVDM.","marker":"(Unterthiner et al. 2018)"}],"fun_headline_variants":["Synthetic fluoroscopy from 400 videos lifts guidewire Dice to 0.61","Two-step diffusion generates labeled fluoroscopy, cutting annotation needs","Scene+action diffusion improves guidewire segmentation with fewer labels","Guidewire Dice +0.09 from diffusion-synthesized videos with sparse labels","SF-VD splits video diffusion for label-efficient guidewire segmentation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The motion-distribution model is trained on only 400 fully annotated 16-frame videos, so the whole approach hinges on cardiac fluoroscopy motion being simple enough—small in-view displacements, stable appearance—that a 2D conditional diffusion model can reproduce heartbeat and breathing motion without temporal artifacts.","fun_headline_variants_meta":{"raw":{"variants":["Synthetic fluoroscopy from 400 videos lifts guidewire Dice to 0.61","Two-step diffusion generates labeled fluoroscopy, cutting annotation needs","Scene+action diffusion improves guidewire segmentation with fewer labels","Guidewire Dice +0.09 from diffusion-synthesized videos with sparse labels","SF-VD splits video diffusion for label-efficient guidewire segmentation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000933,"raw_usage":{"total_tokens":3989,"prompt_tokens":939,"completion_tokens":3050,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":555,"completion_tokens_details":{"reasoning_tokens":2957}},"tokens_in":555,"tokens_out":3050,"duration_ms":18590,"temperature":1.0,"reasoning_tokens":2957,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T10:50:50.322409+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train SF-VD on fluoroscopy videos that violate the simple-motion assumption, such as rapid table panning, large out-of-view device motion, or significant breathing-induced organ displacement, then measure the downstream Dice of a fixed segmentation network trained with the generated videos; if the augmentation no longer improves over no augmentation, the motion-model assumption is the failure point. A second, more direct falsifier is to check whether the frame-consistency sampling actually prevents temporal artifacts by computing FVD on videos generated from held-out mask sequences and asking human raters to detect frame-to-frame anatomical discontinuities.","supporting_citations":[{"cited_title":"F.; Kohl, S","cited_arxiv_id":null,"evidence_quote":"Provides the nnU-Net/Dynvnet architecture used both as a downstream segmentation network and as the backbone for the segmentation-guided mechanism."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies PARN, a wire-specific segmentation network whose performance with SF-VD augmentation is the paper's evidence for benefit on tailored architectures."},{"cited_title":"D.; and Palli, G","cited_arxiv_id":null,"evidence_quote":"Supplies Ariadne+, another wire-specific baseline whose improvement with SF-VD augmentation is reported."}],"review_version":1}