Pith. sign in

REVIEW 4 major objections 4 minor 46 references

An Autoencoder and Vision Transformer-based Interpretability Analysis of the Differences in Automated Staging of Second and Third Molars

T0 review · 4 major / 4 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 The paper claims that the third-molar staging accuracy gap is caused by intrinsic morphological variability in the data, not by the deep learning architecture, and demonstrates a multi-modal interpretability framework that surfaces this.

desk verdict A useful interpretability framework and case study, but the headline claim that tooth 38's poor staging is "morphology, not model" is undercut by an uncontrolled bounding-box confound. read the letter →

arxiv 2509.09911 v1 pith:6UMPGPVH submitted 2025-09-12 cs.CV cs.AI

Barkin Buyukcakir , Jannick De Tobel , Patrick Thevissen , Dirk Vandermeulen , Peter Claes This is my paper · ORCID
classification cs.CVcs.AI MSC 68T0768T4592C55
keywords forensicageestimationdentalstagingthirdmolarvisiontransformerautoencoderinterpretabilitylatentspaceanalysisattentionmaps
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 establish that the large accuracy gap in automated dental staging between mandibular second molars (tooth 37, ~81% accuracy) and third molars (tooth 38, ~54%) is not caused by a weak deep-learning architecture but by properties of the third-molar images themselves: high morphological variability within each developmental stage and weak separation between stages. To show this, it introduces an autoencoder plus Vision Transformer pipeline whose reconstructions, latent-space geometry, and attention maps are all inspected together. The authors argue that attention maps alone are misleading here because they can look anatomically plausible even when the model is failing, and that the added evidence from reconstructions and latent-space distances locates the problem in the data. A sympathetic reader would care because if true, improving forensic age estimation for third molars means improving data collection and stage definitions, not merely swapping models, and it demonstrates a general way to diagnose why a medical AI underperforms.

What carries the argument

The load-bearing object is a convolutional autoencoder trained with a modified triplet margin in which the margin grows with developmental-stage distance, so nearby stages sit close in an L2-normalized latent space and distant stages separate; reconstruction losses (binary cross-entropy plus LPIPS perceptual similarity) keep outputs anatomically readable while nudging each stage toward a prototype. That latent space doubles as the diagnostic instrument: inter-class centroid distances and intra-class pairwise cosine distances quantify separability, and the decoder's reconstructions expose whether stage prototypes are sharp or blurry. The Vision Transformer classifier, with 16 heads and 12 lay

What would settle it

Retrain the same AE+ViT pipeline on tooth 38 images that have been normalized so the tooth occupies the same relative scale as tooth 37 within the cropping window (or vice versa). If accuracy, latent-space separation, and reconstruction sharpness for tooth 38 rise to near tooth 37 levels, the paper's data-centric explanation is confounded by a bounding-box artifact; if they stay at about 0.54 accuracy with poor separation, the conclusion is supported.

Watch

Extended reading notes

Core claim

The paper's central claim, stated in its conclusion, is that 'the poor staging performance on tooth 38 is not an architectural failure of the deep learning models but a data-centric problem rooted in the high intrinsic morphological variability of third molars.' The evidence is produced by a framework that first trains a convolutional autoencoder with an ordinal triplet margin plus BCE and LPIPS reconstruction losses, then trains a Vision Transformer on the reconstructed images as a 10-class staging classifier. Across 4-fold cross-validation the AE+ViT pipeline reaches 0.815 accuracy on tooth 37 and 0.543 on tooth 38, versus 0.712 and 0.462 for the ViT alone. The paper argues that three conv

Load-bearing premise

The paper assumes the larger bounding boxes used for tooth 38—which make the tooth appear relatively smaller than tooth 37—do not themselves cause the lower accuracy or the apparent intra-class variability, since this preprocessing difference is never controlled or tested.

Editorial extensions

If this is right

  • AE preprocessing improves ViT accuracy on both teeth (tooth 37: 0.712 to 0.815; tooth 38: 0.462 to 0.543) while reducing mean absolute error, implying the reconstruction step makes stage features easier to separate.
  • The same AE prefix hurts DenseNet-201 accuracy, suggesting the smoothing is representation-specific and that architecture choice interacts with preprocessing.
  • Plausible attention maps are insufficient for high-stakes medical decisions; they must be supplemented by latent-space and reconstruction diagnostics to reveal data-side failure modes.
  • The tooth 38 dataset's high intra-class and low inter-class variation implies collecting more standardized or better-defined third-molar images could yield more improvement than larger models.
  • The framework can function as a second-opinion tool: on low-confidence cases it can show the expert the reconstruction, nearest latent neighbors, and attention maps.

Reading between the lines

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

  • The paper itself notes in Figure 3 that tooth 38 appears relatively smaller than tooth 37 because of larger bounding boxes; an untested implication is that this preprocessing difference, rather than intrinsic morphology alone, could be inflating the apparent intra-class variability and lowering accuracy.
  • If the data-centric conclusion holds, the same diagnostic pipeline could be applied to other underperforming anatomical sites to decide whether to invest in labels, acquisition protocols, or model capacity.
  • The ordinal triplet latent space could be reused as a confidence score: distance to stage centroids or local intra-class density might flag uncertain predictions for expert review.
  • A natural extension is to register or synthesize third-molar images to reduce morphological variation and check whether accuracy moves toward the tooth 37 level, which would quantify how much of the gap is removable.
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 / 4 minor

Summary. The paper proposes an autoencoder (AE) plus Vision Transformer (ViT) framework for automated dental staging of mandibular second (tooth 37) and third (tooth 38) molars from panoramic radiographs. On a single-site dataset of ~400 images per tooth, the framework improves ViT accuracy from 0.712 to 0.815 for tooth 37 and from 0.462 to 0.543 for tooth 38. Using AE reconstructions, latent-space metrics, and attention maps, the authors argue that the remaining poor performance on tooth 38 is not an architectural failure but a data-centric problem caused by high intrinsic morphological variability of third molars.

Significance. If the central claim is established, the framework would be a useful diagnostic tool for distinguishing model limitations from data limitations in high-stakes forensic age estimation, and the paper's emphasis on multi-faceted interpretability beyond attention maps is valuable. The study is clearly motivated, the experiments are straightforward, and the quantitative accuracy improvements over the ViT baseline are a useful contribution. However, the central explanatory claim depends on an uncontrolled preprocessing difference between the two tooth datasets, and the supporting evidence is not yet conclusive.

major comments (4)
  1. [Methods/Data and Fig. 3] The central claim that tooth 38's lower accuracy is due to intrinsic morphological variability is confounded by the acknowledged difference in bounding boxes. The Fig. 3 caption states that tooth 38 'appears relatively smaller than tooth 37 due to the use of larger bounding boxes.' After resizing to 224×224, tooth 38 occupies fewer pixels, has more background, and has more positional variability. All the evidence used for the data-centric conclusion—lower ViT/DenseNet accuracy, blurrier AE reconstructions (Fig. 6), larger intra-class latent distances (Fig. 11), and crown-focused attention (Fig. 8)—is sensitive to this scale/background mismatch. A smaller, less consistently positioned tooth would plausibly produce the same symptoms even if the true morphological variability were no higher than for tooth 37. The authors should test this directly, e.g., by re-cropping or rescaling tooth 38
  2. [Results and Discussion / Table 1] The main text reports only mean (std) accuracies over four folds and does not provide significance tests or confidence intervals for the key comparisons. In particular, the improvement for tooth 38 (ViT 0.462±0.020 vs. AE+ViT 0.543±0.052) and the tooth 37 vs. tooth 38 performance gap are load-bearing for the central claim. The manuscript mentions 'statistical significance' in Supplementary Table 2, but no such results are presented or summarized in the main text. The authors should report paired tests or confidence intervals for these comparisons, and also for the latent-space metrics (inter/intra-class cosine distances) that are used as quantitative evidence of higher morphological variability.
  3. [Latent Space Analysis] The latent-space evidence is partially circular as evidence of intrinsic data properties. The AE is trained with the stage labels through a triplet loss with an ordinal margin (Eq. 4); the resulting latent space is therefore optimized to separate stages. The observation that tooth 38 embeddings are less separable and more spread out may indicate that the optimizer failed to find a good separation, but it does not independently establish that the data are intrinsically more variable. This is especially true because the AE is trained on the same preprocessed images whose bounding-box properties differ between teeth. The authors should either treat the latent-space results as model-based evidence only, or supplement them with label-free or morphology-based measures (e.g., image registration residuals, shape descriptors, or human-measured anatomical variability) obtained under matched prepro
  4. [Conclusion / Limitations] The conclusion states that the poor staging performance on tooth 38 is 'a data-centric problem rooted in the high intrinsic morphological variability of third molars.' This overstates what the evidence supports. The manuscript's own Synthesis section lists three interpretations: (1) the specific sample collection fails to represent stages consistently, (2) inherent third-molar morphological variability, or (3) a combination. The data presented cannot distinguish between these, and the single-site, ~400-image dataset and acknowledged label variability further limit the claim. The conclusion should be restricted to what the matched and controlled analyses can support.
minor comments (4)
  1. [Fig. 10 caption] The caption refers to 'stages 6 to 10' in the red-rectangle explanation, but the staging range is 0 to 9. This is likely a typo for 'stages 6 to 9.'
  2. [Eq. (4)] The definition of α_ordinal is ambiguous: it is written as ∥y_a − y_n/9∥, which could be read as y_a − (y_n/9) rather than |y_a − y_n|/9. Please clarify the intended scaling.
  3. [Figures] Several figure captions and panel labels contain spacing artifacts (e.g., 'T ooth 37', 'T rue Stage'). These should be cleaned for publication.
  4. [Data and code availability] The data availability statement explains that data cannot be openly shared, but there is no statement about code availability. Providing the preprocessing and training code would strengthen reproducibility, especially because the bounding-box issue is central.

Circularity Check

0 steps flagged · score 1.0 of 10

No load-bearing circularity: held-out accuracies and an independent DenseNet baseline carry the empirical argument; the residual risk is an uncontrolled bounding-box confound, which is a validity concern, not a by-construction reduction.

full rationale

The paper's own derivations are not circular. Classification accuracies (Table 1) are measured on held-out test folds for ViT-only, AE+ViT, DenseNet-only and AE+DenseNet; the gains of AE+ViT over ViT-only and the tooth 37/38 disparity are not fitted parameters renamed as predictions. The disparity is reproduced with two independent architectures (ViT and DenseNet) and agrees with an external preceding study (Matthijs et al. 2024), so the central observation does not rest solely on the authors' own model outputs. The latent-space metrics and reconstructions are used as post-hoc diagnostics, not as fitted inputs: no equation defines 'intra-class morphological variability' in terms of the AE distances, and the conclusion is explicitly hedged (three possible interpretations) before being attributed to third-molar morphology via external literature (refs 33, 34). Self-citations (refs 11, 14, 18; overlapping authors) motivate the hypothesis and supply the staging scheme, but the experiments that verify the hypothesis are new, so the self-citation is not load-bearing. The main threat to the conclusion is a preprocessing confound, not circularity: Fig. 3's caption states that 'Tooth 38 appears relatively smaller than tooth 37 due to the use of larger bounding boxes', and the paper never controls for this scale difference when attributing blurrier reconstructions, larger intra-class cosine distances, and crown-only attention to 'high intrinsic morphological variability'. A smaller tooth with more background and positional variation after resizing would plausibly produce all three symptoms even with equal biological variability. This should be tested with matched crops before the data-centric claim is accepted, but because the claim is supported by measured, externalizable evidence rather than by definitional reduction, the circularity score is low.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

No new physical or conceptual entities are introduced; the AE and ViT are existing architectures. The free parameters listed are model design choices that influence the latent space and therefore the diagnostic interpretation.

free parameters (2)
  • loss weight gamma = 0.7
    Weights triplet loss vs reconstruction loss in Eq. 3; chosen by the authors, not justified by sensitivity analysis.
  • latent dimension = 30 (Eq. 2 says 32)
    Dimension of the AE embedding space; chosen by design; text and Figure 1 are inconsistent.
assumptions (4)
  • domain assumption Expert-assigned Demirjian stages are treated as ground truth labels
    All training and evaluation uses these labels; label noise from inter-observer variability is acknowledged but not modeled.
  • domain assumption Attention rollout reflects the ViT's decision-relevant focus
    Used as the interpretability tool despite cited work questioning whether attention equals explanation.
  • domain assumption The AE latent space with L2-normalized embeddings and triplet loss gives a meaningful ordinal metric
    The diagnostic conclusions rely on distances in this space.
  • domain assumption The single-institution Belgian dataset is representative for comparing the two teeth
    The paper notes this limitation; all conclusions are drawn from this sample.

how reviews work

0 comments
Cite this review

Pith. "Pith review of An Autoencoder and Vision Transformer-based Interpretability Analysis of the Differences in Automated Staging of Second and Third Molars." pith.science (2026). https://pith.science/paper/6UMPGPVH

@misc{pith2026250909911,
  author       = {Pith},
  title        = {Pith review of: An Autoencoder and Vision Transformer-based Interpretability Analysis of the Differences in Automated Staging of Second and Third Molars},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6UMPGPVH}},
  note         = {Machine review of arXiv:2509.09911}
}
read the original abstract

The practical adoption of deep learning in high-stakes forensic applications, such as dental age estimation, is often limited by the 'black box' nature of the models. This study introduces a framework designed to enhance both performance and transparency in this context. We use a notable performance disparity in the automated staging of mandibular second (tooth 37) and third (tooth 38) molars as a case study. The proposed framework, which combines a convolutional autoencoder (AE) with a Vision Transformer (ViT), improves classification accuracy for both teeth over a baseline ViT, increasing from 0.712 to 0.815 for tooth 37 and from 0.462 to 0.543 for tooth 38. Beyond improving performance, the framework provides multi-faceted diagnostic insights. Analysis of the AE's latent space metrics and image reconstructions indicates that the remaining performance gap is data-centric, suggesting high intra-class morphological variability in the tooth 38 dataset is a primary limiting factor. This work highlights the insufficiency of relying on a single mode of interpretability, such as attention maps, which can appear anatomically plausible yet fail to identify underlying data issues. By offering a methodology that both enhances accuracy and provides evidence for why a model may be uncertain, this framework serves as a more robust tool to support expert decision-making in forensic age estimation.

Figures

Figures reproduced from arXiv: 2509.09911 by the authors.

Figure 1
Figure 1. Overview of the complete pipeline. There are two sequential trainings taking place, the first being of the autoencoder (top) and the second of the ViT (bottom) on the image reconstructions. The autoencoder consists of a symmetric encoder-decoder setup of three layers each, with linear projections to and from the embedding space. The final layer of the decoder ends with a sigmoid activation in order to limit the pixe… view at source ↗
Figure 2
Figure 2. The number of samples per stage in the two datasets. The samples are balanced in stage and in sex, eliminating the need for methods which rectify class imbalance, and facilitating sex-invariant training. validation set. Each split was stratified, preserving the balanced distribution of stage and sex of the full datasets. DenseNet-201 To establish a performance baseline for our study, a pretrained DenseNet-201 model,… view at source ↗
Figure 3
Figure 3. Schematic representations and illustrative case examples of the developmental stages, based on the modified Demirjian staging technique. Note the subtle differences in the bounding box sizes surrounding teeth 37 and 38. Tooth 38 appears relatively smaller than tooth 37 due to the use of larger bounding boxes, which were deliberately sized to fully encompass each tooth. This approach accounts for the greater anatomic… view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: The violin plots for illustrating the distribution of predicted dental stages for two different models, the DenseNet-201 only (orange) and AE + DenseNet (blue). The analysis is presented separately for tooth 37 and 38. For tooth 37, both models show a strong correleati…
Figure 5
Figure 5. Figure 5: The distributions of ViT stage predictions versus ground truth labels per stage, from the AE + ViT framework and using the ViT model only without image reconstruction. The predictions are acquired via cross-validation. It can be seen that the predictions for tooth 37 a…
Figure 6
Figure 6. Figure 6: Examples of original samples and AE reconstructions, contrasted against mean stage images for tooth 37 and tooth 38. Tooth 37 reconstructions show a denoising effect while reframing images as stage prototypes, and offer readable reconstructions which are sharper than t…
Figure 7
Figure 7. Figure 7: Comparison of the mean stage images and attention maps for tooth 37. Top row: Mean stage images, a general representation of stage shape. Second row: Mean ViT attention maps, showing how the ViT focus changes across stages by moving from the tooth cusp to the cervical …
Figure 8
Figure 8. Figure 8: Comparison of mean stage images and attention maps for tooth 38. Top row: Mean stage images reveal a blurrier average per stage, indicating the tooth shapes for 38 contain more intra-class variation. Second row: Mean ViT attention maps seem similar to those of tooth 37…
Figure 9
Figure 9. Figure 9: Heatmaps of LPIPS loss between the attention maps of each sample in the test set, per fold and model. The samples are ordered by stage label, and axis labels denote changes of label. Each fold is visualized independently. The LPIPS loss signifies perceptual similarity …
Figure 10
Figure 10. Figure 10: The PCA plots of latent space representations of the test set images, for both teeth from all folds. The embeddings are L2-normalized to mimic their contribution in the triplet loss during training. Dashed rectangles are utilised to highligh areas of high overlap of d…
Figure 11
Figure 11. Figure 11: An analysis of the latent space grouping of the embeddings for both teeth. The embeddings from the test sets of all folds were used for the calculations. Top row shows the inter-class distances, defined as the pairwise cosine distance between the L2-normalized geometr…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

46 extracted references · 1 canonical work pages

  1. [1]

    & Van Dillen, L

    Gummerum, M., López-Pérez, B., Van Dijk, E. & Van Dillen, L. F. When punishment is emotion-driven: Children’s, adolescents’, and adults’ costly punishment of unfair allocations.Soc. Dev.29, 126–142 (2020)

    2020

  2. [2]

    & Geserick, G

    Schmeling, A., Dettmeyer, R., Rudolf, E., Vieth, V . & Geserick, G. Forensic age estimation: methods, certainty, and the law.Deutsches Ärzteblatt Int.113, 44 (2016)

    2016

  3. [3]

    & Yoshimura, H

    Matsuda, S., Yoshida, H., Ebata, K., Shimada, I. & Yoshimura, H. Forensic odontology with digital technologies: A systematic review.J. forensic legal medicine74, 102004 (2020)

    2020

  4. [4]

    Lewis, J. M. & Senn, D. R. Forensic dental age estimation: an overview.J. California Dental Assoc.43, 315–319 (2015)

    2015

  5. [5]

    W., Fieuws, S

    Thevissen, P. W., Fieuws, S. & Willems, G. Third molar development: evaluation of nine tooth development registration techniques for age estimations.J. forensic sciences58, 393–397 (2013)

    2013

  6. [6]

    Niu, Z., Zhong, G. & Yu, H. A review on the attention mechanism of deep learning.Neurocomputing452, 48–62 (2021)

    2021

  7. [7]

    A., Davies, J

    Rahim, A. A., Davies, J. & Liversidge, H. Reliability and limitations of permanent tooth staging techniques.Forensic Sci. Int.346, 111654 (2023)

    2023

  8. [8]

    & Liversidge, H

    Dhanjal, K., Bhardwaj, M. & Liversidge, H. Reproducibility of radiographic stage assessment of third molars.Forensic science international159, S74–S77 (2006)

    2006

Show all 46 references
  1. [9]

    P., Nilendu, D., Thomas, N., Nagpal, S

    Pillai, J. P., Nilendu, D., Thomas, N., Nagpal, S. & Nedunari, L. S. S. Inter-observer agreement in the radiographic interpretation of demirjian’s developmental stages in the mandibular second and third molars–a comparative study.J. Oral Maxillofac. Pathol.25, 554–555 (2021)

    2021

  2. [10]

    M.et al.Medical image analysis using convolutional neural networks: a review.J

    Anwar, S. M.et al.Medical image analysis using convolutional neural networks: a review.J. medical systems42, 1–13 (2018)

    2018

  3. [11]

    & Thevissen, P

    De Tobel, J., Radesh, P., Vandermeulen, D. & Thevissen, P. W. An automated technique to stage lower third molar development on panoramic radiographs for age estimation: a pilot study.The J. forensic odonto-stomatology35, 42 (2017)

    2017

  4. [12]

    Banar, N.et al.Towards fully automated third molar development staging in panoramic radiographs.Int. J. Leg. Medicine 134, 1831–1841 (2020)

    2020

  5. [13]

    Han, M.et al.With or without human interference for precise age estimation based on machine learning?Int. J. Leg. Medicine136, 821–831 (2022)

    2022

  6. [14]

    Matthijs, L.et al.Artificial intelligence and dental age estimation: development and validation of an automated stage allocation technique on all mandibular tooth types in panoramic radiographs.Int. J. Leg. Medicine138, 2469–2479 (2024)

    2024

  7. [15]

    & Lagioia, F

    Sartor, G. & Lagioia, F. The impact of the general data protection regulation (gdpr) on artificial intelligence. InPanel for the Future of Science and Technology(2020). 16.Kroll, J. A.Accountable algorithms. Ph.D. thesis, Princeton University (2015). 17.Danks, D. & London, A. ...

    2020

  8. [18]

    Forensic Sci.(2024)

    Büyükçakır, B.et al.Opg-based dental age estimation using a data-technical exploration of deep learning techniques.J. Forensic Sci.(2024)

    2024

  9. [19]

    National Science and Technology Council (US). Select Committee on Artificial Intelligence.The National Artificial Intelligence Research and Development Strategic Plan: 2023 Update(National Science and Technology Council (US), Select Committee on Artificial . . . , 2019)

    2023

  10. [20]

    T., Dosovitskiy, A., Brox, T

    Springenberg, J. T., Dosovitskiy, A., Brox, T. & Riedmiller, M. Striving for simplicity: The all convolutional net.arXiv preprint arXiv:1412.6806(2014)

    2014 arXiv

  11. [21]

    R.et al.Grad-cam: Visual explanations from deep networks via gradient-based localization

    Selvaraju, R. R.et al.Grad-cam: Visual explanations from deep networks via gradient-based localization. InProceedings of the IEEE international conference on computer vision, 618–626 (2017)

    2017

  12. [22]

    & Marcu, A

    Yasin, N., Hare, J. & Marcu, A. Is saliency really captured by gradient? InNeurIPS 2024 Workshop on Scientific Methods for Understanding Deep Learning(2024)

    2024

  13. [23]

    & Wand, M

    Woerl, A.-C., Disselhoff, J. & Wand, M. Initialization noise in image gradients and saliency maps. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 1766–1775 (2023). 24.Vaswani, A.et al.Attention is all you need.Adv. neural information proce...

    2023

  14. [25]

    Dosovitskiy, A.et al.An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929(2020)

    2010 arXiv

  15. [26]

    & Bernardino, J

    Maurício, J., Domingues, I. & Bernardino, J. Comparing vision transformers and convolutional neural networks for image classification: A literature review.Appl. Sci.13, 5521 (2023)

    2023

  16. [27]

    why should i trust you?

    Ribeiro, M. T., Singh, S. & Guestrin, C. " why should i trust you?" explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, 1135–1144 (2016)

    2016

  17. [28]

    AI Lett.2, e51 (2021)

    Ray, A.et al.Generating and evaluating explanations of attended and error-inducing input regions for vqa models.Appl. AI Lett.2, e51 (2021). 29.Jain, S. & Wallace, B. C. Attention is not explanation.arXiv preprint arXiv:1902.10186(2019). 30.Wiegreffe, S. & Pinter, Y . Attentio...

    2021 arXiv

  18. [31]

    Saporta, A.et al.Deep learning saliency maps do not accurately highlight diagnostically relevant regions for medical image interpretation.MedRxiv(2021)

    2021

  19. [32]

    Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead.Nat

    Rudin, C. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead.Nat. machine intelligence1, 206–215 (2019)

    2019

  20. [33]

    A., Bani Younis, H

    Al-Qudah, A. A., Bani Younis, H. A. B., Awawdeh, L. A. & Daud, A. Root and canal morphology of third molar teeth.Sci. reports13, 6901 (2023)

    2023

  21. [34]

    & Ohshima, H

    Morita, W., Morimoto, N. & Ohshima, H. Exploring metameric variation in human molars: a morphological study using morphometric mapping.J. anatomy229, 343–355 (2016)

    2016

  22. [35]

    & Weinberger, K

    Huang, G., Liu, Z., Van Der Maaten, L. & Weinberger, K. Q. Densely connected convolutional networks. InProceedings of the IEEE conference on computer vision and pattern recognition, 4700–4708 (2017)

    2017

  23. [36]

    & Willems, G

    Thevissen, P., Fieuws, S. & Willems, G. Human dental age estimation using third molar developmental stages: does a bayesian approach outperform regression models to discriminate between juveniles and adults?Int. journal legal medicine 124, 35–42 (2010)

    2010

  24. [37]

    H., Vieira, S., Garcia-Dias, R

    Lopez Pinaya, W. H., Vieira, S., Garcia-Dias, R. & Mechelli, A. Chapter 11 - autoencoders. In Mechelli, A. & Vieira, S. (eds.)Machine Learning, 193–208, DOI: https://doi.org/10.1016/B978-0-12-815739-8.00011-0 (Academic Press, 2020)

    2020 doi

  25. [38]

    & Huang, K

    Chen, W., Chen, X., Zhang, J. & Huang, K. Beyond triplet loss: a deep quadruplet network for person re-identification. In Proceedings of the IEEE conference on computer vision and pattern recognition, 403–412 (2017)

    2017

  26. [39]

    & Zheng, N

    Cheng, D., Gong, Y ., Zhou, S., Wang, J. & Zheng, N. Person re-identification by multi-channel parts-based cnn with improved triplet loss function. InProceedings of the iEEE conference on computer vision and pattern recognition, 1335–1344 (2016)

    2016

  27. [40]

    A., Shechtman, E

    Zhang, R., Isola, P., Efros, A. A., Shechtman, E. & Wang, O. The unreasonable effectiveness of deep features as a perceptual metric. InProceedings of the IEEE conference on computer vision and pattern recognition, 586–595 (2018). 20/21

    2018

  28. [41]

    & Lee, H

    Chung, Y . & Lee, H. Joint triplet loss with semi-hard constraint for data augmentation and disease prediction using gene expression data.Sci. Reports13, 18178 (2023). 42.Liversidge, H. M. Timing of human mandibular third molar formation.Annals human biology35, 294–321 (2008)....

    2023 arXiv

  29. [44]

    BMC Oral Heal.24, 426 (2024)

    Ong, S.-H.et al.Fully automated deep learning approach to dental development assessment in panoramic radiographs. BMC Oral Heal.24, 426 (2024)

    2024

  30. [45]

    Maia, M. C. G., Martins, M. d. G. A., Germano, F. A., Neto, J. B. & Da Silva, C. A. B. Demirjian’s system for estimating the dental age of northeastern brazilian children.Forensic science international200, 177–e1 (2010)

    2010

  31. [46]

    & Brki´c, H

    Ambarkova, V ., Gali´c, I., V odanovi´c, M., Bioˇcina-Lukenda, D. & Brki´c, H. Dental age estimation using demirjian and willems methods: cross sectional study on children from the former yugoslav republic of macedonia.Forensic science international234, 187–e1 (2014)

    2014

  32. [47]

    development of a reference dataset and comparison with a united kingdom caucasian reference dataset.J

    Elshehawi, W.et al.Dental age assessment of maltese children and adolescents. development of a reference dataset and comparison with a united kingdom caucasian reference dataset.J. Forensic Leg. Medicine39, 27–33 (2016)

    2016

  33. [48]

    Merdietio Boedi, R.et al.Evaluation of demirjian staging method reproducibility and reliability for dental age estimation between eight forensic experts.Int. J. Leg. Medicine1–8 (2025)

    2025

  34. [49]

    J., Petrie, A

    Boonpitaksathit, T., Hunt, N., Roberts, G. J., Petrie, A. & Lucas, V . S. Dental age assessment of adolescents and emerging adults in united kingdom caucasians using censored data for stage h of third molar roots.The Eur. J. Orthod.33, 503–508 (2011)

    2011

  35. [50]

    Franco, A., Vetter, F., Coimbra, E. d. F., Fernandes, Â. & Thevissen, P. Comparing third molar root development staging in panoramic radiography, extracted teeth, and cone beam computed tomography.Int. journal legal medicine134, 347–353 (2020)

    2020

  36. [51]

    forensic odonto-stomatology35, 117 (2017)

    De Tobel, J.et al.Forensic age estimation based on development of third molars: a staging technique for magnetic resonance imaging.The J. forensic odonto-stomatology35, 117 (2017)

    2017

  37. [52]

    journal legal medicine133, 583–592 (2019)

    De Tobel, J.et al.Magnetic resonance imaging of third molars in forensic age estimation: comparison of the ghent and graz protocols focusing on apical closure.Int. journal legal medicine133, 583–592 (2019)

    2019

  38. [53]

    InSeminars in Musculoskeletal Radiology, vol

    De Tobel, J.et al.Dental and skeletal imaging in forensic age estimation: disparities in current approaches and the continuing search for optimization. InSeminars in Musculoskeletal Radiology, vol. 24, 510–522 (Thieme Medical Publishers, 2020). Acknowledgements This work was f...

    2020

Pith tools

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