Pith. sign in

REVIEW 4 major objections 5 minor 21 references

Tooth-Diffusion: Guided 3D CBCT Synthesis with Fine-Grained Tooth Conditioning

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A conditional diffusion model makes individual teeth appear or disappear in 3D dental scans.

desk verdict New per-tooth conditioning for 3D CBCT synthesis, but the evaluation never checks that the generated dentition actually matches the condition vector. read the letter →

arxiv 2508.14276 v1 pith:HAZCHS7Z submitted 2025-08-19 cs.CV cs.AI

Said Djafar Said , Torkan Gholamalizadeh , Mostafa Mehdipour Ghazi This is my paper
classification cs.CVcs.AI
keywords CBCTsynthesistoothconditioningdiffusionmodelswaveletFiLMdentalimaging3Dinpaintingmedicalimagegeneration
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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

What would settle it

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.

Watch

Extended reading notes

Core claim

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

Load-bearing premise

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.

Editorial extensions

If this is right

  • 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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • 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.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

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.

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 (4)
  1. [§4.3, Table 1, abstract] 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).
  2. [§4.2, §4.3, Fig. 2, Table 1] 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.
  3. [§4.3] 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.
  4. [§4.1] 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.
minor comments (5)
  1. [§3.4, Eq. (2)] The masked-loss weight λ is set to 10 with no sensitivity analysis or ablation. A small sweep would help justify this choice.
  2. [§4.2] 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.
  3. [Fig. 2] 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.
  4. [§4.1] 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.
  5. [§5] 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.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the conditional diffusion model is trained and tested on external CBCT data, and matched train/test task protocols are standard supervised evaluation, not derivation-by-construction.

full rationale

I walked the paper's derivation chain and found no load-bearing circular step. The method is an empirical conditional diffusion model: a binary tooth-presence vector is an input condition, the model is trained with a masked L2 loss plus a wavelet-diffusion loss, and it is evaluated on held-out CBCT volumes. No self-definitional step appears: the tooth-presence vector is not defined in terms of the generated volumes; it comes from manual/segmentation-derived labels. No fitted parameter is renamed as a prediction: hyperparameters such as λ=10 are fixed before evaluation, and the test metrics are computed on unseen patient scans rather than on the training set. The tooth-addition and tooth-removal test tasks mirror the training augmentation protocol, but this is a standard supervised evaluation design for conditional generation, not a prediction statistically forced by construction—the test set contains unseen patients, and no equation reduces the SSIM/FID numbers to the training loss. The paper contains no self-citations that carry the argument; the dataset and backbone (WDM, FiLM) are cited from external prior work. No uniqueness theorem is imported from the authors' own prior work, and no known result is renamed. The skeptical concern that no direct metric verifies condition adherence is an evaluation-validity caveat, not circularity: FID and SSIM compare against real volumes and ground-truth regions, not against the conditioning input itself. I therefore find no circular step that can be quoted and reduced, and assign score 0.

Assumptions & free parameters 1 free parameters · 5 assumptions · 0 invented entities

The central claim rests on the accuracy of manual relabeling, the realism of the inpainting augmentation, the sufficiency of binary presence vectors, and the validity of 3D FID with an unspecified feature extractor. No invented entities are introduced. The single hand-set free parameter is the masked loss weight lambda=10.

free parameters (1)
  • masked loss weight lambda = 10
    Set empirically in Eq. 2; chosen by hand, not derived.
assumptions (5)
  • domain assumption Manual relabeling of 98 CBCT volumes to Universal Numbering System is anatomically accurate
    Central to ground truth conditions and evaluation; no inter-observer agreement reported.
  • domain assumption Distance-transform inpainting produces realistic missing-tooth anatomy for training
    Used in Section 3.3 to simulate missing teeth; if artifacts remain, the model may learn unrealistic patterns.
  • domain assumption 32-bit tooth presence vector is a sufficient condition for anatomically realistic tooth generation
    The paper assumes this conditioning signal captures the required anatomical variability.
  • domain assumption 3D FID with Inception features is a valid distributional metric for CBCT volumes
    No details on how Inception features are obtained for 3D volumes; metric validity assumed.
  • standard math Standard DDPM/WDM denoising assumptions hold for CBCT intensities normalized to [-1,1]
    Inherited from [9,18].

how reviews work

0 comments
Cite this review

Pith. "Pith review of Tooth-Diffusion: Guided 3D CBCT Synthesis with Fine-Grained Tooth Conditioning." pith.science (2026). https://pith.science/paper/HAZCHS7Z

@misc{pith2026250814276,
  author       = {Pith},
  title        = {Pith review of: Tooth-Diffusion: Guided 3D CBCT Synthesis with Fine-Grained Tooth Conditioning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HAZCHS7Z}},
  note         = {Machine review of arXiv:2508.14276}
}
read the original abstract

Despite the growing importance of dental CBCT scans for diagnosis and treatment planning, generating anatomically realistic scans with fine-grained control remains a challenge in medical image synthesis. In this work, we propose a novel conditional diffusion framework for 3D dental volume generation, guided by tooth-level binary attributes that allow precise control over tooth presence and configuration. Our approach integrates wavelet-based denoising diffusion, FiLM conditioning, and masked loss functions to focus learning on relevant anatomical structures. We evaluate the model across diverse tasks, such as tooth addition, removal, and full dentition synthesis, using both paired and distributional similarity metrics. Results show strong fidelity and generalization with low FID scores, robust inpainting performance, and SSIM values above 0.91 even on unseen scans. By enabling realistic, localized modification of dentition without rescanning, this work opens opportunities for surgical planning, patient communication, and targeted data augmentation in dental AI workflows. The codes are available at: https://github.com/djafar1/tooth-diffusion.

Figures

Figures reproduced from arXiv: 2508.14276 by the authors.

Figure 1
Figure 1. Overview of the proposed framework. A guided diffusion model is used [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Similarities between the original and reconstructed tooth when individ [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Qualitative comparison between generated CBCT scans and their corre [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

21 extracted references · 20 canonical work pages

  1. [1]

    Scarfe, W.C., Farman, A.G.: What is cone-beam CT and how does it work? Dental Clinics of North America52(4) (2008) 707–730

    work page 2008

  2. [2]

    DentomaxillofacialRadiology 44(1)(2015)20140224

    Pauwels, R., Araki, K., Siewerdsen, J., Thongvigitmanee, S.S.: Technical aspects of dentalCBCT:stateoftheart. DentomaxillofacialRadiology 44(1)(2015)20140224

    work page 2015

  3. [3]

    Dentomaxillofacial Radiology53(1) (2024) 5–21

    Sadr, S., Rokhshad, R., Daghighi, Y., Golkar, M., Tolooie Kheybari, F., Gorjinejad, F., Mataji Kojori, A., Rahimirad, P., Shobeiri, P., Mahdian, M., et al.: Deep learningfortoothidentificationandnumberingondentalradiography:asystematic review and meta-analysis. Dentomaxillofacial Radiology53(1) (2024) 5–21

    work page 2024

  4. [4]

    Expert Systems with Applications199 (2022)

    Singh, N.K., Raza, K.: Progress in deep learning-based dental and maxillofacial image analysis: A systematic review. Expert Systems with Applications199 (2022)

    work page 2022

  5. [5]

    In: Proceedings of the Computer Vision and Pattern Recognition Con- ference

    Bolelli, F., Marchesini, K., van Nistelrooij, N., Lumetti, L., Pipoli, V., Ficarra, E., Vinayahalingam, S., Grana, C.: Segmenting maxillofacial structures in CBCT volumes. In: Proceedings of the Computer Vision and Pattern Recognition Con- ference. (2025) 5238–5248

    work page 2025

  6. [6]

    Scientific Reports 13(1) (2023) 6031 4 https://ditto.ing.unimore.it/toothfairy3/ 10 S

    Choi, H., Yun, J.P., Lee, A., Han, S.S., Kim, S.W., Lee, C.: Deep learning synthe- sis of cone-beam computed tomography from zero echo time magnetic resonance imaging. Scientific Reports 13(1) (2023) 6031 4 https://ditto.ing.unimore.it/toothfairy3/ 10 S. D. Said et al

    work page 2023

  7. [7]

    Scientific Reports 15(1) (2025) 10816

    Hu, C., Cao, N., Li, X., He, Y., Zhou, H.: CBCT-to-CT synthesis using a hybrid U-Net diffusion model based on transformers and information bottleneck theory. Scientific Reports 15(1) (2025) 10816

    work page 2025

  8. [8]

    Medical Image Analysis99 (2025) 103362

    Zhang, Y., Li, L., Wang, J., Yang, X., Zhou, H., He, J., Xie, Y., Jiang, Y., Sun, W., Zhang, X., et al.: Texture-preserving diffusion model for CBCT-to-CT synthesis. Medical Image Analysis99 (2025) 103362

    work page 2025

Show all 21 references
  1. [9]

    Advances in Neural Information Processing Systems33 (2020) 6840–6851

    Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems33 (2020) 6840–6851

    2020

  2. [10]

    arXiv:2506.13545 (2025)

    Gao, Y., Pan, S., Hu, M., Xie, H., Remick, J., Chang, C.W., Roper, J., Tian, Z., Yang, X.: Limited-angle CBCT reconstruction via geometry-integrated cycle- domain denoising diffusion probabilistic models. arXiv:2506.13545 (2025)

    2025 arXiv

  3. [11]

    arXiv preprint arXiv:2504.00264 (2025)

    Demir, B., Liu, Y., Chen, X., Chen, E.Z., Zhao, L., Mailhe, B., Chen, T., Sun, S.: DiffDenoise: self-supervised medical image denoising with conditional diffusion models. arXiv preprint arXiv:2504.00264 (2025)

    2025 arXiv

  4. [12]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Perez, E., Strub, F., De Vries, H., Dumoulin, V., Courville, A.: FiLM: Visual rea- soning with a general conditioning layer. In: Proceedings of the AAAI Conference on Artificial Intelligence. Volume 32. (2018)

    2018

  5. [13]

    Clinical Oral Investigations27(7) (2023) 3363–3378

    Polizzi, A., Quinzi, V., Ronsivalle, V., Venezia, P., Santonocito, S., Lo Giudice, A., Leonardi, R., Isola, G.: Tooth automatic segmentation from CBCT images: a systematic review. Clinical Oral Investigations27(7) (2023) 3363–3378

    2023

  6. [14]

    Journal of Dentistry147 (2024) 105130

    Dot, G., Chaurasia, A., Dubois, G., Savoldelli, C., Haghighat, S., Azimian, S., Taramsari, A.R., Sivaramakrishnan, G., Issa, J., Dubey, A., et al.: DentalSegmen- tator: robust open source deep learning-based CT and CBCT image segmentation. Journal of Dentistry147 (2024) 105130

    2024

  7. [15]

    Journal of Imaging 11(2) (2025) 41

    Pedersen, S., Jain, S., Chavez, M., Ladehoff, V., de Freitas, B.N., Pauwels, R.: Pano-GAN: A deep generative model for panoramic dental radiographs. Journal of Imaging 11(2) (2025) 41

    2025

  8. [16]

    In: International Conference on Machine Learning, PMLR (2017) 214–223

    Arjovsky,M.,Chintala,S.,Bottou,L.: Wassersteingenerativeadversarialnetworks. In: International Conference on Machine Learning, PMLR (2017) 214–223

    2017

  9. [17]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. (2022) 10684–10695

    2022

  10. [18]

    In: MICCAI Work- shop on Deep Generative Models, Springer (2024) 11–21

    Friedrich, P., Wolleb, J., Bieder, F., Durrer, A., Cattin, P.C.: WDM: 3D wavelet diffusion models for high-resolution medical image synthesis. In: MICCAI Work- shop on Deep Generative Models, Springer (2024) 11–21

    2024

  11. [19]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Cui, Z., Li, C., Wang, W.: ToothNet: automatic tooth instance segmentation and identification from cone beam CT images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. (2019) 6368–6377

    2019

  12. [20]

    In: Information Processing in Medical Imaging: 27th International Conference, Virtual Event, June 28–June 30, 2021, Proceedings 27, Springer (2021) 150–162

    Cui, Z., Zhang, B., Lian, C., Li, C., Yang, L., Wang, W., Zhu, M., Shen, D.: Hierarchical morphology-guided tooth instance segmentation from CBCT images. In: Information Processing in Medical Imaging: 27th International Conference, Virtual Event, June 28–June 30, 2021, Proceed...

    2021

  13. [21]

    Nature Communications13(1) (2022)

    Cui, Z., Fang, Y., Mei, L., Zhang, B., Yu, B., Liu, J., Jiang, C., Sun, Y., Ma, L., Huang, J., et al.: A fully automatic AI system for tooth and alveolar bone segmentation from cone-beam CT images. Nature Communications13(1) (2022)

    2022

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.