Pith. sign in

REVIEW 4 major objections 5 minor 40 references

Multimodal Contrastive Pretraining of CBCT and IOS for Enhanced Tooth Segmentation

T0 review · 4 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read This paper introduces ToothMCL, a multimodal contrastive pretraining framework that aligns CBCT and intraoral-scan representations so a single pretrained encoder improves multi-class tooth segmentation under FDI numbering across internal an

desk verdict A useful dataset and a controlled ablation support a modest pretraining benefit, but the headline gains are overstated, the cross-modal matching rule is undefined, and the paper's own tables contradict its third-molar claims. read the letter →

arxiv 2509.07923 v1 pith:TLKZ4FWB submitted 2025-09-09 cs.CV cs.AI

Moo Hyun Son , Juyoung Bae , Zelin Qiu , Jiale Peng , Kai Xin Li , Yifan Lin , Hao Chen This is my paper · ORCID
classification cs.CVcs.AI
keywords toothsegmentationmultimodalcontrastivelearningCBCTintraoralscanself-supervisedpretrainingpointcloudFDInumberingdentalimaging
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

The paper tries to show that pretraining on paired cone-beam CT (CBCT) and intraoral scan (IOS) data teaches a model anatomical features that transfer to supervised tooth segmentation. It introduces ToothMCL, a patch-level contrastive method that pulls CBCT and IOS representations together, and CBCT-IOS3.8K, the largest paired dataset of its kind with 3,867 patients. On five public benchmarks the pretrained models beat single-modality baselines, with DSC gains of 12 percentage points for CBCT and 8 for IOS on out-of-distribution data. If this holds, dental AI can build one reusable backbone from routinely collected paired scans instead of training separate models per modality.

What carries the argument

The load-bearing mechanism is a dual patch-level alignment objective. Two encoders—a Swin Transformer for CBCT volumes and a Point Transformer for IOS point clouds—produce patch embeddings; a sigmoid cross-modal contrastive loss (Eq. 4) pulls matched CBCT-IOS patches together and pushes unmatched ones apart, while an intra-modal loss (Eq. 8) pulls patches that depict similar anatomy within each modality. The two are combined as L_total = L_intra + α L_cross, and the pretrained encoders initialize separate supervised fine-tuning decoders for CBCT and IOS segmentation.

What would settle it

Train the identical pipeline on CBCT-IOS pairs shuffled across patients so no true cross-modal correspondence exists; if downstream Dice gains stay near the reported 2.7–9.9 percentage points, the cross-modal alignment loss is not the operative mechanism. Registering a held-out pair and checking whether the peak-similarity patches correspond to the same anatomical teeth would also settle it.

Watch

Extended reading notes

Core claim

The central claim is that multimodal contrastive pretraining over paired volumetric and surface dental scans produces modality-invariant, anatomy-aware representations that transfer to multi-class tooth segmentation with FDI numbering. The authors report state-of-the-art results: 93.38% DSC on internal CBCT, 88.92% on external CBCT, 93.16–96.49% DSC on IOS datasets, and consistent gains across all tooth groups and clinically difficult cases such as impacted teeth, metal artifacts, and crossbite. The contribution is framed as the first dental multimodal pretraining framework plus a large paired dataset.

Load-bearing premise

The central assumption is that a CBCT patch and an IOS patch that count as a positive pair actually show the same tooth in the same place; the paper never states how patches are paired, so the alignment could be learning patient identity instead of anatomy.

Editorial extensions

If this is right

  • Pretrained features raise external CBCT segmentation by a 12 percentage-point DSC margin over the next-best baseline and keep every tooth group above 90% DSC.
  • Pretrained features raise external IOS segmentation by an 8 percentage-point DSC margin over the next-best baseline, with consistently stronger per-tooth performance except for third molars.
  • Pretraining gains scale monotonically with paired-data size, with no saturation through 3,867 patients, so collecting more paired scans should continue to help.
  • The same pretrained encoders transfer across scanners, voxel spacings, and patient populations, and handle horizontally impacted teeth, metal artifacts, and severe malocclusion.

Reading between the lines

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

  • Because the paper does not specify how CBCT and IOS patches are paired as positives, the learned invariance could be patient identity rather than anatomical correspondence; a shuffled-pair control would separate those explanations.
  • If the matching rule is anatomical, the same pretraining should transfer to other paired dental modalities, such as panoramic and cephalometric X-rays, which the paper notes are out of scope.
  • The unified latent space could serve as a shared backbone for downstream tasks beyond segmentation, including caries detection, orthodontic simulation, and prosthesis design.
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 ToothMCL, a self-supervised multimodal contrastive pretraining framework that aligns CBCT volumetric patches with IOS point-cloud patches, and introduces CBCT-IOS3.8K, a private paired dataset of 3,867 patients. After pretraining, modality-specific encoders (SwinUNETR for CBCT, Point Transformer for IOS) are fine-tuned for multi-class tooth segmentation with FDI numbering, and evaluated on five public benchmarks: ToothFairy2 and Cui for CBCT; Teeth3DS, TADPM, and 3D-IOSSeg for IOS. The authors report state-of-the-art DSC on internal and external tests, with claimed relative improvements of 12% for CBCT and 8% for IOS, and present with/without-pretraining ablations and pretraining-scale curves as evidence that the pretraining is the active ingredient.

Significance. If the results hold, the contribution is substantial: it is the first pretraining framework using paired CBCT/IOS data, it provides a large private paired dataset, and it evaluates on five public benchmarks with statistical testing and qualitative clinical case analysis. The with/without-pretraining comparison in Table A4 is a useful causal control. However, the central mechanism—cross-modal patch alignment—is not operationally defined, and the ablation text reports numbers that disagree with Table A4. The significance of the work therefore depends on resolving these issues; the current version does not make the claimed cross-modal alignment mechanism testable.

major comments (4)
  1. [Method: CBCT-IOS Multimodal Contrastive Learning, Eqs. (4)-(9)] The positive-pair rule for the central contrastive loss is never defined. Eq. (4) uses z_ij = 1 for 'matched CBCT and IOS patch pairs' and Eq. (8) uses z_ij = 1 for patches that 'depict similar anatomical structures,' but the paper only defines patient-level pairing D = {(C_m, P_m)} and forms per-scan patch sets c_m = {c_{m,1},...,c_{m,r}}, p_m = {p_{m,1},...,p_{m,s}}. The preprocessing section describes resampling, HU clipping, ROI cropping, and FPS, but no registration, projection, or spatial indexing that could establish patch-level correspondences between a 3D volume and a surface scan. As written, LCross can only be interpreted as a bag-of-patches / patient-identity objective, not the claimed 'fine-grained anatomical features' alignment. This is load-bearing: the paper's novelty is precisely the cross-modal patch alignment, and Eqs. (4)-(9) are not computable without the missing rul
  2. [Results: Ablation Study, Fig. 7 and Table A4] The ablation paragraph states that pretraining improves overall DSC by 2.7-9.9 pp, with the smallest gain on ToothFairy2 at 2.69 pp and the largest on Cui at 9.85 pp. These numbers do not match Table A4. From Table A4, the differences are: ToothFairy2 93.38-88.28 = 5.10 pp, Cui 88.92-85.83 = 3.09 pp, Teeth3DS 93.16-87.86 = 5.30 pp, TADPM 96.49-95.85 = 0.64 pp, and 3D-IOSSeg 93.60-90.01 = 3.59 pp. The actual range is 0.64-5.30 pp, and the smallest gain is on TADPM, not ToothFairy2. The reported 2.69 pp and 9.85 pp match comparisons against SSL in Table A2, not the with/without-pretraining control in Table A4. This inconsistency undermines the specific quantitative claims in the ablation and must be corrected.
  3. [Results: Ablation Study] The paper does not ablate the components of the pretraining objective. The only controlled comparison is full ToothMCL pretraining vs. random initialization (Table A4). Since the contribution is specifically the cross-modal loss LCross (Eq. 4) and the dual intra-modal/cross-modal design, an ablation with LIntra only, LCross only, or a unimodal pretraining baseline is needed to show that the cross-modal alignment, rather than generic self-supervised representation learning on a large dataset, is responsible for the gains. Without this, a reader cannot attribute the observed improvements to the proposed multimodal alignment mechanism.
  4. [Evaluation protocol / Data Availability] The paper never specifies the train/validation/test splits for the five public datasets, nor does it define what makes a dataset 'internal' vs. 'external' beyond naming. The Data Availability section states that 'The public dataset was used for training and evaluating the downstream task,' but it is not stated which datasets were used for fine-tuning and which only for testing. This is important for the external-generalization claim: if, for example, all public datasets were used for training in some way, the external-testing interpretation changes. Please provide explicit split protocols and dataset usage details.
minor comments (5)
  1. [Abstract / Results (CBCT and IOS external evaluation)] The 'increase of 12% for CBCT segmentation and 8% for IOS segmentation' is not labeled as relative vs. absolute. From Table A2, the external CBCT comparison with the next-best method is 88.92 vs. 79.07 = 9.85 absolute points, which is 12% relative; from Table A3, the IOS external comparison is 93.60 vs. 86.86 = 6.74 absolute points, which is 8% relative. Please state the baseline and the type of percentage to avoid ambiguity.
  2. [Method, Eqs. (3)-(5)] The notation is inconsistent: p_m is defined in Eq. (3) as the encoded IOS patch representations, but Eq. (5) redefines p_m as the encoded patch set from a generic modality-specific encoder f(P_m). This makes it unclear whether intra-modal alignment is being applied to both modalities and how the two modalities share the same variable name.
  3. [Method, Eq. (8)] The intra-modal loss formula is difficult to parse. The sums over m and n, the role of w, and the max over i are not explained in the text. Please clarify the index ranges and the intuition behind this formulation.
  4. [Results, Pretraining scale (Fig. 7b)] The pretraining-scale experiment reports only textual increments (+3.79 pp from 25% to 50%, +5.87 pp from 75% to 100%) with no table or exact values for the subsampled conditions. Reporting the underlying DSC values and confidence intervals would make the monotonicity claim verifiable.
  5. [Results, IOS external evaluation (3D-IOSSeg, Table A3)] The text says the model performed best 'except in the third molar group' but does not say which method wins there. Table A3 shows TGNet achieves 63.24 vs. Ours 55.22 for the third molar on 3D-IOSSeg; this should be acknowledged explicitly.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: pretraining gains are benchmarked on independent public datasets with a from-scratch control; undefined patch matching is an ambiguity, not a circular reduction.

full rationale

ToothMCL's reported gains are not reduction-by-construction. The pretraining objective (Eqs. 4, 8, 9) is an input to training, not a fitted quantity renamed as a result. Downstream segmentation is fine-tuned on public labeled datasets (ToothFairy2, Cui, Teeth3DS, TADPM, 3D-IOSSeg) and compared against independent baselines, with the from-scratch control in Table A4 providing an external anchor; the 12% and 8% margins are measured on external sets against published methods. No load-bearing self-citation chain or imported uniqueness theorem is used. The only weakness worth flagging is that the patch-matching rule for z_ij in Eqs. 4 and 8 is not operationally specified (no registration, projection, or spatial indexing is described), which is a methodological ambiguity/correctness concern rather than circularity: even a patient-identity match would still supply a real training signal, and the downstream evaluation would not be forced by it. Therefore no circular step meets the evidentiary bar.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on (1) four learnable hyperparameters in the contrastive losses, (2) an undefined patch-correspondence rule that makes the cross-modal and intra-modal positives meaningful, and (3) a pairing assumption for the private dataset (near-contemporaneous CBCT and IOS per patient). The evaluation is against external public benchmarks, so the main risk is not circularity but underspecification and internal inconsistency.

free parameters (5)
  • alpha (cross-modal loss weight, Eq 9) = not reported
    Balances L_intra and L_cross in L_total; no value or sensitivity analysis is given.
  • t_g, b_g (cross-modal logit scale and bias, Eq 4) = not reported
    Learned parameters of the sigmoid contrastive loss; values never reported.
  • t_l, b_l (intra-modal logit scale and bias, Eq 7) = not reported
    Learned parameters of the intra-modal loss; values never reported.
  • Patch geometry (patch counts r, s, hidden dim h, patch size) = not reported
    Defines the granularity of the alignment losses; none of these numbers are given, and no ablation over them is presented.
  • Pretraining batch size = set by GPU memory
    'Batch size determined by available GPU memory capacity'; not fixed or reported per run, which affects the contrastive objective directly.
assumptions (4)
  • ad hoc to paper Cross-modal positive pairs in Eq 4 correspond to the same anatomical region in CBCT and IOS
    Introduced without definition in 'CBCT-IOS Multimodal Contrastive Learning' (Eqs 4-9). No registration or spatial matching rule is described; the 'fine-grained anatomical' mechanism depends on it.
  • ad hoc to paper Intra-modal positives ('patches depict similar anatomical structures', Eq 8) are identifiable in self-supervised data
    The pretraining is unlabeled, yet Eq 7 requires binary labels z_ij for patch pairs. The source of these labels (augmentation, geometry, or something else) is never stated.
  • domain assumption CBCT and IOS scans in CBCT-IOS3.8K are anatomically comparable per patient (near-contemporaneous capture)
    Data were collected 2022-2024 from patients aged 15-86; no temporal alignment between the two modalities is reported. Growing or moved teeth would weaken the pairing. Invoked by the entire pretraining stage.
  • domain assumption Contrastive pretraining on unlabeled paired data transfers to supervised segmentation on other public datasets
    Standard transfer premise; supported empirically only by Table A4, whose interpretation in the text is internally inconsistent (see red flags).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multimodal Contrastive Pretraining of CBCT and IOS for Enhanced Tooth Segmentation." pith.science (2026). https://pith.science/paper/TLKZ4FWB

@misc{pith2026250907923,
  author       = {Pith},
  title        = {Pith review of: Multimodal Contrastive Pretraining of CBCT and IOS for Enhanced Tooth Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TLKZ4FWB}},
  note         = {Machine review of arXiv:2509.07923}
}
read the original abstract

Digital dentistry represents a transformative shift in modern dental practice. The foundational step in this transformation is the accurate digital representation of the patient's dentition, which is obtained from segmented Cone-Beam Computed Tomography (CBCT) and Intraoral Scans (IOS). Despite the growing interest in digital dental technologies, existing segmentation methodologies frequently lack rigorous validation and demonstrate limited performance and clinical applicability. To the best of our knowledge, this is the first work to introduce a multimodal pretraining framework for tooth segmentation. We present ToothMCL, a Tooth Multimodal Contrastive Learning for pretraining that integrates volumetric (CBCT) and surface-based (IOS) modalities. By capturing modality-invariant representations through multimodal contrastive learning, our approach effectively models fine-grained anatomical features, enabling precise multi-class segmentation and accurate identification of F\'ed\'eration Dentaire Internationale (FDI) tooth numbering. Along with the framework, we curated CBCT-IOS3.8K, the largest paired CBCT and IOS dataset to date, comprising 3,867 patients. We then evaluated ToothMCL on a comprehensive collection of independent datasets, representing the largest and most diverse evaluation to date. Our method achieves state-of-the-art performance in both internal and external testing, with an increase of 12\% for CBCT segmentation and 8\% for IOS segmentation in the Dice Similarity Coefficient (DSC). Furthermore, ToothMCL consistently surpasses existing approaches in tooth groups and demonstrates robust generalizability across varying imaging conditions and clinical scenarios.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

40 extracted references · 39 canonical work pages

  1. [1]

    Global oral health status report: Towards universal health coverage for oral health by 2030 (2022)

    World Health Organization. Global oral health status report: Towards universal health coverage for oral health by 2030 (2022)

    work page

  2. [2]

    C., Farman, A

    Scarfe, W. C., Farman, A. G. & Sukovi´ c, M. Clinical applications of cone-beam computed tomography in dental practice.Journal of the Canadian Dental Association72, 75–80 (2006)

    work page 2006

  3. [3]

    & Logozzo, S

    Mangano, F., Gandolfi, A., Luongo, G. & Logozzo, S. Intraoral scanners in dentistry: a review of the current literature.BMC Oral Health17, 149 (2017)

    work page 2017

  4. [4]

    B., Fontenele, R

    Baldini, B., Papasratorn, D., Fagundes, F. B., Fontenele, R. C. & Jacobs, R. Validation of a novel tool for automated tooth modelling by fusion of cbct-derived roots with the respective ios-derived crowns.Journal of Dentistry153, 105546 (2025)

    work page 2025

  5. [5]

    Albano, D.et al.Artificial intelligence for radiographic imaging detection of caries lesions: a systematic review.BMC Oral Health24, 274 (2024)

    work page 2024

  6. [6]

    Scientific Reports14, 26752 (2024)

    Adnan, N.et al.Developing an ai-based application for caries index detection on intraoral photographs. Scientific Reports14, 26752 (2024)

    work page 2024

  7. [7]

    Negi, S.et al.Artificial intelligence in dental caries diagnosis and detection: An umbrella review.Clinical and Experimental Dental Research10, e70004 (2024)

    work page 2024

  8. [8]

    & Shan, Z

    Liu, J., Zhang, C. & Shan, Z. Application of artificial intelligence in orthodontics: Current state and future perspectives.Healthcare11, 2760 (2023)

    work page 2023

Show all 40 references
  1. [9]

    Deng, Q., Yang, X., Huang, M., Jiang, L. & Zhang, D.Taposenet: Teeth alignment based on pose estima- tion via multi-scale graph convolutional network.Proceedings of Medical Image Computing and Computer Assisted Intervention – MICCAI 2024, Vol. 15012, 314–323 (Springer Nature S...

    2024

  2. [10]

    Lei, C.et al.Automatic tooth arrangement with joint features of point and mesh representations via diffusion probabilistic models.Computer Aided Geometric Design111, 102293 (2024)

    2024

  3. [11]

    & Kim, Y.-L

    Kong, H.-J. & Kim, Y.-L. Application of artificial intelligence in dental crown prosthesis: A scoping review. BMC Oral Health24, 937 (2024)

    2024

  4. [12]

    & Keren, J.From mesh completion to ai designed crown.Proceedings of Medical Image Computing and Computer-Assisted Intervention – MICCAI 2023, Vol

    Hosseinimanesh, G., Ghadiri, F., Guibault, F., Cheriet, F. & Keren, J.From mesh completion to ai designed crown.Proceedings of Medical Image Computing and Computer-Assisted Intervention – MICCAI 2023, Vol. 12905 ofLecture Notes in Computer Science, 543–552 (Springer, 2023)

    2023

  5. [13]

    Chau, R. C. W., Hsung, R. T.-C., Koohi-Moghadam, M. & Lam, W. Y. H. Accuracy of artificial intelligence- designed single-molar dental prostheses: A feasibility study.The Journal of Prosthetic Dentistry131, 1111–1117 (2024)

    2024

  6. [14]

    Chen, X., Ma, N., Xu, T. & Xu, C. Deep learning-based tooth segmentation methods in medical imaging: A review.Proceedings of the Institution of Mechanical Engineers, Part H: Journal of Engineering in Medicine 238, 3–20 (2024)

    2024

  7. [15]

    & Haffner, P

    LeCun, Y., Bottou, L., Bengio, Y. & Haffner, P. Gradient-based learning applied to document recognition. Proceedings of the IEEE86, 2278–2324 (1998)

    1998

  8. [16]

    Ronneberger, O., Fischer, P. & Brox, T.U-net: Convolutional networks for biomedical image segmenta- tion.Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015, 234–241 (Springer International Publishing, 2015)

    2015

  9. [17]

    Cui, Z.et al.A fully automatic ai system for tooth and alveolar bone segmentation from cone-beam ct images.Nature Communications13, 2096 (2022)

  10. [18]

    Attention is all you need.Advances in Neural Information Processing Systems, 5998–6008 (2017)

    Vaswani, A.et al. Attention is all you need.Advances in Neural Information Processing Systems, 5998–6008 (2017). 15

    2017

  11. [19]

    Gillot, M.et al.Automatic multi-anatomical skull structure segmentation of cone-beam computed tomography scans using 3d unetr.PLOS ONE17, e0275033 (2022)

    2022

  12. [20]

    & Dao, T.Mamba: Linear-time sequence modeling with selective state spaces.Proceedings of the Conference on Language Modeling (COLM)(2025)

    Gu, A. & Dao, T.Mamba: Linear-time sequence modeling with selective state spaces.Proceedings of the Conference on Language Modeling (COLM)(2025)

    2025

  13. [21]

    Hao, J.et al.T-mamba: A unified framework with long-range dependency in dual-domain for 2d & 3d tooth segmentation.arXiv preprint arXiv:2404.01065(2024)

    2024 arXiv

  14. [22]

    C., Hagenbuchner, M

    Scarselli, F., Gori, M., Tsoi, A. C., Hagenbuchner, M. & Monfardini, G. The graph neural network model. IEEE Transactions on Neural Networks20, 61–80 (2009)

    2009

  15. [23]

    & Mahjoub, M

    Ben-Hamadou, A., Smaoui, O., Chaabouni-Chouayakh, H. & Mahjoub, M. A. Teeth3ds: A benchmark for teeth segmentation and labeling from intra-oral 3d scans.arXiv preprint arXiv:2210.06094(2022)

    2022

  16. [24]

    Dentistry—Designation system for teeth and areas of the oral cavity

    International Organization for Standardization. Dentistry—Designation system for teeth and areas of the oral cavity. ISO 3950:2016 (2016)

    2016

  17. [25]

    & Krois, J

    Schwendicke, F., Samek, W. & Krois, J. Artificial intelligence in dentistry: Chances and challenges.Journal of Dental Research99, 769–774 (2020)

    2020

  18. [26]

    & Hinton, G.A simple framework for contrastive learning of visual representations.Proceedings of the 37th International Conference on Machine Learning (ICML), 1597–1607 (2020)

    Chen, T., Kornblith, S., Norouzi, M. & Hinton, G.A simple framework for contrastive learning of visual representations.Proceedings of the 37th International Conference on Machine Learning (ICML), 1597–1607 (2020)

    2020

  19. [27]

    Learning transferable visual models from natural language supervision.Proceedings of the 38th International Conference on Machine Learning (ICML)(2021)

    Radford, A.et al. Learning transferable visual models from natural language supervision.Proceedings of the 38th International Conference on Machine Learning (ICML)(2021)

    2021

  20. [28]

    Tang, Y.et al. Self-supervised pre-training of swin transformers for 3d medical image analysis.Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 20730–20740 (2022)

    2022

  21. [29]

    F., Kohl, S

    Isensee, F., Jaeger, P. F., Kohl, S. A. A., Petersen, J. & Maier-Hein, K. H. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation.Nature Methods18, 203–211 (2021)

    2021

  22. [30]

    R., Su, H., Mo, K

    Qi, C. R., Su, H., Mo, K. & Guibas, L. J.Pointnet: Deep learning on point sets for 3d classification and segmentation.Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 652–660 (2017)

    2017

  23. [31]

    Wang, Y.et al.Dynamic graph cnn for learning on point clouds.ACM Transactions on Graphics (TOG) (2019)

    2019

  24. [32]

    Tsegformer: 3d tooth segmentation in intraoral scans with geometry guided transformer

    Xiong, H.et al. Tsegformer: 3d tooth segmentation in intraoral scans with geometry guided transformer. Proceedings of Medical Image Computing and Computer Assisted Intervention – MICCAI 2023(2023)

    2023

  25. [33]

    IEEE Transactions on Medical Imaging1–17 (2024)

    Bolelli, F.et al.Segmenting the Inferior Alveolar Canal in CBCTs Volumes: the ToothFairy Challenge. IEEE Transactions on Medical Imaging1–17 (2024)

    2024

  26. [34]

    Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images.Brainlesion: Glioma, Multiple Sclerosis, Stroke and Traumatic Brain Injuries, Vol

    Hatamizadeh, A.et al. Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images.Brainlesion: Glioma, Multiple Sclerosis, Stroke and Traumatic Brain Injuries, Vol. 12962 ofLecture Notes in Computer Science, 272–284 (2022)

    2022

  27. [35]

    Pytorch: An imperative style, high-performance deep learning library.Advances in Neural Information Processing Systems 32, 8024–8035 (2019)

    Paszke, A.et al. Pytorch: An imperative style, high-performance deep learning library.Advances in Neural Information Processing Systems 32, 8024–8035 (2019)

    2019

  28. [36]

    & Hutter, F.Decoupled weight decay regularization.International Conference on Learning Representations(2019)

    Loshchilov, I. & Hutter, F.Decoupled weight decay regularization.International Conference on Learning Representations(2019)

    2019

  29. [37]

    & Hutter, F.SGDR: Stochastic gradient descent with warm restarts.International Conference on Learning Representations (ICLR)(2017)

    Loshchilov, I. & Hutter, F.SGDR: Stochastic gradient descent with warm restarts.International Conference on Learning Representations (ICLR)(2017)

    2017

  30. [38]

    Zhao, H., Jiang, L., Jia, J., Torr, P. H. S. & Koltun, V.Point transformer.Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 16259–16268 (2021). 16

    2021

  31. [39]

    Li, J.et al.A fine-grained orthodontics segmentation model for 3d intraoral scan data.Computers in Biology and Medicine168, 107821 (2024)

    2024

  32. [40]

    Cipriano, M., Allegretti, S., Bolelli, F., Pollastri, F. & Grana, C.Improving Segmentation of the Inferior Alveolar Nerve through Deep Label Propagation.IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 21105–21114 (IEEE, 2022). 17 Appendix A Extended Data...

    2022

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.