{"id":"92c1ba11-a142-4b15-be75-d57a604a0533","arxiv_id":"2508.14276","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A conditional wavelet diffusion model with FiLM conditioning generates and edits 3D CBCT volumes from per-tooth presence vectors.","lead":"A diffusion model that generates and edits 3D dental CBCT scans based on which teeth are present, tested for tooth addition, removal, and full-mouth synthesis. The approach could support surgical planning and data augmentation in dental imaging, but the evaluation uses a very small, non-random test set.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No metric verifies that generated volumes actually satisfy the requested tooth-presence vector; FID/SSIM similarity cannot by itself establish fine-grained control.","rationale":"The reader's weakest assumption (8-patient test set) is real but secondary. The more load-bearing gap is that the headline property—conditioned tooth presence—is never measured directly. The reported metrics can be high even if the model does not actually follow the condition vector: full-volume FID is dominated by background and global intensity statistics, while SSIM is computed on the tooth region after the fact and does not test whether the condition caused the tooth to appear or disappear. This is not an internal inconsistency or a fundamental flaw: the method is plausible, code and data are released, and a segmentation-based condition-adherence evaluation would settle it. Because the gap is fillable with additional experiments rather than requiring a redesign, the reader's CONDITIONAL verdict remains the right calibration. I partially agree with the reader's concern: we both identify evaluation insufficiency, but I would pinpoint the lack of direct condition-adherence verification as the weakest assumption, with the small test set as a compounding factor. The stray path-like text between Fig. 2 and Table 1 does not affect this assessment.","tokens_in":7561,"tokens_out":4695,"duration_ms":58545,"concrete_test":"Use the provided ground-truth tooth segmentation protocol to segment generated volumes. For each of the 8 test patients, sample at least 10 volumes per condition (e.g., add tooth 1, remove teeth [16,17], full dentition, empty dentition), then compute per-tooth presence precision/recall against the requested binary vector, per-tooth Dice between generated and ground-truth tooth regions for added teeth, and false-positive voxel rates for removed teeth, stratified by tooth ID. Also report 95% confidence intervals across patients and sampling seeds. If presence accuracy is not essentially perfect and removed teeth are not suppressed (false-positive rate near zero), the fine-grained-control claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is fine-grained control over individual tooth presence, but the paper reports no quantitative test of condition adherence. In Section 4.3, tooth addition is scored by SSIM/PSNR between the output and the original unaltered tooth, and tooth removal is scored by full-volume FID between generated scans and real scans with matching absence patterns. Both metrics are similarity-based and computed over at most 8 test patients. A model that produces realistic but wrong dentition—e.g., regenerates a partially similar or neighboring tooth, or fails to actually suppress the specified tooth—can still achieve high SSIM/FID because FID over 256^3 background-dominated volumes is insensitive to localized tooth-level configuration, and SSIM is computed only after aligning to the ground-truth target tooth. Thus “precise control over tooth presence” (abstract, Section 1) is not directly established: the causal link between the binary condition vector and the resulting dentition is exactly what is unevidenced. The 8-patient test set compounds this, since per-tooth numbers in Fig. 2 rest on at most 8 samples per tooth, with no error bars or condition-adherence rates reported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Tooth-Diffusion, a conditional diffusion framework for 3D CBCT dental volume synthesis. The model is conditioned on a 32-dimensional binary tooth-presence vector via FiLM embeddings, uses wavelet-domain denoising, and is trained with an additional masked L2 loss that emphasizes tooth regions. Training includes simulated tooth-addition and tooth-removal augmentations. The authors evaluate reconstruction, tooth addition, tooth removal, and full-dentition synthesis on a dataset of 98 CBCT volumes, holding out 8 edge-case test scans, using 3D FID and SSIM/PSNR. The central claim is fine-grained, user-controllable tooth presence in 3D CBCT generation, enabling clinical simulations and data augmentation.","tokens_in":7826,"tokens_out":5508,"duration_ms":65332,"significance":"If established, the ability to realistically add or remove individual teeth in 3D CBCT volumes without rescanning would be valuable for surgical planning, patient communication, and targeted data augmentation in dental AI. The paper addresses a relevant and under-explored problem, releases code, and combines reasonable components (wavelet diffusion, FiLM conditioning, masked loss) in a clinically oriented setup. However, the current evidence base is too thin to support the central claim of precise control: there is no direct condition-adherence metric, no baselines or ablations, no error bars or statistical analysis, and the patient-level split is not clearly documented. The paper is a plausible proof-of-concept but the quantitative evaluation needs substantial strengthening before the stated conclusions are justified.","major_comments":[{"comment":"The central claim is precise control over tooth presence, but no experiment directly verifies that the generated volume actually satisfies the conditioning vector. Tooth addition is scored by SSIM/PSNR against the ground-truth target tooth, and tooth removal by FID against real volumes with matching absence patterns. Both are similarity-based and can be insensitive to localized tooth-level errors: a model that ignores the condition and produces a plausible but incorrect dentition could still score well. I request a direct condition-adherence evaluation (e.g., per-tooth presence/absence accuracy, Dice overlap between generated and target tooth masks, or control experiments showing that changing only one bit of the condition vector changes only the corresponding tooth region).","section":"§4.3, Table 1, abstract"},{"comment":"All quantitative claims rest on 8 test scans, yet no error bars, confidence intervals, or statistical tests are reported. Per-tooth averages in Fig. 2 are based on at most 8 samples and fewer for rare teeth, so the per-tooth fidelity comparisons are not stable. The 3D FID computation is also underspecified: no feature extractor is named, and the number of volumes used for the reference set is not given. Since FID is known to be sensitive to small sample sizes (also acknowledged by the validation set of 2 scans in §4.3), the claim of 'strong generalization' is not supported by the reported numbers alone.","section":"§4.2, §4.3, Fig. 2, Table 1"},{"comment":"No baseline or ablation is provided. The contributions listed in §1 (FiLM conditioning, masked L2 loss, wavelet diffusion, augmentation) are not isolated or compared against a simpler alternative, such as an unconditional WDM, a FiLM-free conditional diffusion, or a standard 3D U-Net inpainting model. Without such comparisons, the reported fidelity cannot be attributed to the proposed components, and the 'novelty' of the approach is not empirically demonstrated.","section":"§4.3"},{"comment":"The train/test split is not defined at the patient level. The text says '8 unique patient scans' are reserved for testing, but the dataset contains patients with multiple CBCT acquisitions at different treatment stages (§4.1). If the same patient appears in both training and test, the generalization results would be inflated by patient-level leakage. Please explicitly confirm that all 90 training/validation scans come from patients disjoint from the 8 test patients, and if not, rerun the evaluation on a properly disjoint split.","section":"§4.1"}],"minor_comments":[{"comment":"The masked-loss weight λ is set to 10 with no sensitivity analysis or ablation. A small sweep would help justify this choice.","section":"§3.4, Eq. (2)"},{"comment":"The statement that the higher validation FID (88.81) is 'attributed to the smaller number of samples (2 vs. 8)' itself illustrates the instability of FID at these sample sizes. Please report the exact numbers of generated and real volumes used in every FID computation.","section":"§4.2"},{"comment":"The figure lacks error bars and confidence intervals, and the tooth IDs are not clearly readable. The caption says 'when individually removed and regenerated' although the experiment is tooth addition (masking a tooth and reconstructing it); please clarify the terminology.","section":"Fig. 2"},{"comment":"The phrase 'For missing teeth annotations, we provide manual annotations where applicable' is ambiguous: are missing teeth labeled as absent, or are missing annotations manually completed? Please clarify.","section":"§4.1"},{"comment":"The conclusion states 'high visual fidelity and generalization' but no clinical or user study is reported. Please temper the conclusion to the actual evidence, or include such an evaluation.","section":"§5"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a worthwhile problem and the code release is a plus, but the evaluation is not yet at the level required for the central claim of fine-grained control. The most important concerns are the absence of a direct condition-adherence metric and the lack of patient-level split documentation. I would like to see the authors add a concrete verification that the conditioning vector is causally reflected in the output, together with baselines and basic uncertainty quantification, before this is ready for acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The contribution is real but the evaluation is too thin to back the central claim. The paper builds a conditional diffusion model for 3D CBCT where a 32-bit tooth-presence vector drives addition/removal via FiLM and a masked loss. That combination is new for dental imaging, and the code/data being public is a plus. But nothing in the experiments verifies that the generated volume actually matches the requested tooth-presence vector. SSIM/PSNR are computed only on the target tooth after the fact, and FID over full 256^3 volumes is insensitive to localized tooth configuration. So the fine-grained control claim is not directly evidenced.\n\nThe stress-test note lands. The method is not circular—the training augmentation and eval tasks are consistent—but the evidence is thin: 8 test scans, no baselines, no ablations, no error bars, and a 3D FID whose feature extractor is unspecified. The reconstruction experiment is essentially autoencoding, so a FID of 40.27 doesn't show generalization. The per-tooth results in Fig. 2 rest on at most 8 samples per tooth.\n\nWhat's genuinely good: the idea is sensible, the inpainting strategy for filling tooth cavities before training is thoughtful, and the authors are honest about dataset limitations and about wisdom teeth being hard. That keeps the paper from being a wash.\n\nThis is a niche paper, for dental imaging and medical synthesis people. It deserves a serious referee—the method is new and testable—but a referee should ask for a direct condition-adherence check (e.g., automatically segmenting generated volumes and comparing to the condition), plus baselines, ablations, and error bars. Until then, the only defensible claim is 'we trained it and it produces plausible-looking volumes,' not 'precise control.'","headline":"New per-tooth conditioning for 3D CBCT synthesis, but the evaluation never checks that the generated dentition actually matches the condition vector.","tokens_in":8294,"tokens_out":2536,"would_cite":false,"duration_ms":27636,"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":"A conditional diffusion model makes individual teeth appear or disappear in 3D dental scans.","keywords":["CBCT synthesis","tooth conditioning","diffusion models","wavelet diffusion","FiLM conditioning","dental imaging","3D inpainting","medical image generation"],"falsifier":"Take any held-out CBCT volume not among the 98 and run the tooth-addition task with ground-truth labels; if average per-tooth SSIM drops well below 0.91 or FID moves far outside the reported 40 to 89 range, the claimed generalization fails.","tokens_in":7489,"feed_emoji":"🦷","tokens_out":6044,"duration_ms":61227,"temperature":0.7,"pith_summary":"This paper aims to show that a conditional diffusion model can generate 3D dental cone-beam CT (CBCT) volumes in which the presence or absence of each individual tooth is set by a 32-entry binary code. If the claim holds, clinicians could simulate tooth removal, addition, or a complete dentition directly on an existing patient scan without rescanning, and AI dental models could be trained on targeted synthetic variations. The model runs denoising in a wavelet domain, uses FiLM conditioning to inject the tooth-presence code, and applies a masked L2 loss so learning focuses on tooth regions rather than background. On the paper's test set, reconstruction achieves an FID of 40.27 and full-dentition synthesis reaches SSIM above 0.91 on unseen scans, with per-tooth fidelity lower for wisdom teeth. The contribution is the fine-grained tooth-level control itself, which earlier CBCT generation methods lacked.","feed_headline":"Diffusion model edits individual teeth in 3D dental scans","feed_subtitle":"Simulates tooth removal, addition, and full dentitions from a single patient scan, no rescanning needed.","key_machinery":"The load-bearing object is the 32-dimensional tooth-presence vector, embedded through a linear layer and injected into every residual block via FiLM (feature-wise linear modulation), which applies learned per-condition scaling and shifting to intermediate U-Net activations. Denoising is performed in a 3D Haar wavelet domain (WDM) at half spatial resolution to cut memory and compute. Training augments the conditioning with simulated tooth addition and removal, where cavities are filled using a Manhattan distance transform that propagates nearest valid voxel intensities followed by Gaussian smoothing, and a masked L2 loss with weight 10 concentrates the objective on Gaussian-blurred tooth regi","core_discovery":"The central discovery is that a compact, discrete per-tooth attribute vector is enough to drive anatomically plausible, local edits of a 3D CBCT volume. Given a conditioning image and a binary vector of 32 tooth states, the same trained model performs three operations: it reconstructs a scan with its original dentition, it inpaints teeth that were masked out (addition), and it suppresses teeth that are marked absent (removal), including generating a full dental set from an edentulous scan. The paper reports that this transfers to test patients it never trained on, including edge cases with brackets, braces, and mini-screws, while per-tooth similarity analysis shows the weakest results on the","pith_inferences":["Extending the 32-bit code to include implant, crown, and bridge states would preserve the same architecture; the paper names these as future work.","Because the reported 0.91 SSIM is computed over inpainted regions only, whole-volume comparisons would show lower numbers; evaluations across methods should match the region of interest.","A direct test of augmentation value would be training a tooth-segmentation network on synthetic mixtures of real and generated volumes and measuring Dice gains; the paper leaves that to future work.","The same conditioning mechanism could transfer to neighboring volumetric anatomy, such as maxillofacial bones, by redefining the attribute vector."],"forward_implications":["A single trained model can flip between tooth completion, removal, and full dentition synthesis by changing the binary code at inference time.","Targeted missing-tooth patterns can be generated from full-dentition scans, providing training data for segmentation and detection without new patient radiation.","Treatment scenarios such as implant placement can be visualized on a specific patient's own CBCT before any procedure.","The per-tooth similarity results identify which tooth positions are data-starved, such as molars and wisdom teeth, pointing data collection toward those classes.","Wavelet-domain generation at half resolution makes the approach more feasible on clinical GPUs than pixel-space volumetric diffusion."],"supporting_citations":[{"why":"Supplies the wavelet-domain 3D diffusion backbone that this method adapts to CBCT volumes.","marker":"[18]"},{"why":"Defines the FiLM conditioning layer used to inject the tooth-presence code into the network.","marker":"[12]"},{"why":"Provide the CBCT volumes and ground-truth tooth segmentations used for training and evaluation.","marker":"[19,20,21]"},{"why":"Is the denoising diffusion probabilistic model formulation the approach extends with conditioning.","marker":"[9]"}],"fun_headline_variants":["Tooth-Diffusion: add, remove, or regenerate teeth in 3D CBCT scans","Per-tooth attributes guide realistic 3D dental volume edits","Diffusion model enables localized tooth edits without rescanning","CBCT synthesis with fine-grained tooth conditioning via diffusion"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The evaluation rests on eight test scans chosen as edge cases, and if those eight are not representative of the broader patient population, the reported fidelity and generalization numbers do not establish the claim.","fun_headline_variants_meta":{"raw":{"variants":["Tooth-Diffusion: add, remove, or regenerate teeth in 3D CBCT scans","Per-tooth attributes guide realistic 3D dental volume edits","Diffusion model enables localized tooth edits without rescanning","CBCT synthesis with fine-grained tooth conditioning via diffusion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00028,"raw_usage":{"total_tokens":1485,"prompt_tokens":717,"completion_tokens":768,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":461,"completion_tokens_details":{"reasoning_tokens":692}},"tokens_in":461,"tokens_out":768,"duration_ms":7981,"temperature":1.0,"reasoning_tokens":692,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T18:38:50.916645+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take any held-out CBCT volume not among the 98 and run the tooth-addition task with ground-truth labels; if average per-tooth SSIM drops well below 0.91 or FID moves far outside the reported 40 to 89 range, the claimed generalization fails.","supporting_citations":[{"cited_title":"In: MICCAI Work- shop on Deep Generative Models, Springer (2024) 11–21","cited_arxiv_id":null,"evidence_quote":"Supplies the wavelet-domain 3D diffusion backbone that this method adapts to CBCT volumes."},{"cited_title":"In: Proceedings of the AAAI Conference on Artificial Intelligence","cited_arxiv_id":null,"evidence_quote":"Defines the FiLM conditioning layer used to inject the tooth-presence code into the network."},{"cited_title":"Advances in Neural Information Processing Systems33 (2020) 6840–6851","cited_arxiv_id":null,"evidence_quote":"Is the denoising diffusion probabilistic model formulation the approach extends with conditioning."}],"review_version":1}