Pith. sign in

REVIEW 4 major objections 3 minor 43 references

Attention Maps in 3D Shape Classification for Dental Stage Estimation with Class Node Graph Attention Networks

T0 review · 4 major / 3 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 CGAT stages 3D third molars and returns attention maps that trace each call to crown, furcation, roots, and apices, matching expert staging criteria while reaching a 0.76 weighted F1.

desk verdict A useful architectural variant (GATv2 + CLS node) with a decent classification result, but the explainability claim rests on unvalidated attention rollout and a leaky split. read the letter →

arxiv 2509.07581 v1 pith:YJSQFETS submitted 2025-09-09 cs.CV cs.AI

Barkin Buyukcakir , Rocharles Cavalcante Fontenele , Reinhilde Jacobs , Jannick De Tobel , Patrick Thevissen , Dirk Vandermeulen , Peter Claes This is my paper
classification cs.CVcs.AI
keywords explainableAIgraphattentionnetworks3DshapeclassificationdentalstageassessmentrolloutCLSnodecone-beamCTDemirjianstaging
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 tries to show that a graph neural network can classify 3D tooth shapes into dental developmental stages and, at the same time, show where on the tooth its attention is focused, in a way that matches how forensic dentists read these teeth. The stakes are practical: automated dental staging could remove inter-observer variability in forensic age estimation, but only if clinicians can see why the model made its call. The authors propose the Class Node Graph Attention Network (CGAT), which adds a global CLS node to a graph attention convolution and classifies from that node's embedding, then rolls attention out to each mesh node to produce a map. They test local mean curvature, distance to centroid, and the combination on 528 third-molar meshes spanning Demirjian stages D-H. Their evidence indicates that the combined-feature directed-edge model reaches a 0.76 weighted F1 score, and that attention maps shift from crown to furcation to roots to apices as development progresses, matching expert criteria.

What carries the argument

The load-bearing mechanism is the CLS node: a virtual node appended to every input mesh graph, connected to all mesh nodes (by either directed or undirected edges), whose final embedding alone feeds the classification head. Because the graph attention convolution computes learnable attention weights for every edge into the CLS node, and because all decision-relevant information must pass through that node, those weights—composed over layers with attention rollout—are read as per-node contributions to the decision. A second mechanism is the node feature design: pose-invariant local mean curvature and distance to centroid encode shape in a way that lets the attention maps track the anatomical

What would settle it

Run a trained CGAT on a stage-H tooth, then surgically mask or erase the regions its attention map marks highest (root apices and furcation) and re-run: if the predicted stage stays roughly the same, or if erasing low-attention regions changes the prediction more, the attention map does not reflect the true decision basis.

Watch

Extended reading notes

Core claim

The central claim is that funneling classification through a virtual CLS node in a graph attention network creates attention maps that legitimately explain the model's decision on a 3D mesh. The classification head consumes only the CLS node's final embedding, so every piece of information used for the decision must have flowed through the CLS node's attention to the mesh nodes; the authors then use attention rollout over the learned attention weights to visualize, on the tooth surface, how much each node influenced the CLS embedding. The paper's finding is that these maps align with the anatomical criteria of the Demirjian staging system: early stages focus on the crown, intermediate stages

Load-bearing premise

The attention weights from the CLS node to each mesh node—after attention rollout—are assumed to show how much each node actually contributed to the classification decision.

Editorial extensions

If this is right

  • If attention maps are faithful, clinicians can validate each automated dental-stage call anatomically, directly addressing the trust barrier that has limited deep learning in forensic age estimation.
  • Using directed edges to the CLS node, rather than undirected edges, preserves multi-region attention in deeper models, so interpretability should be treated as a design axis alongside accuracy.
  • Combining curvature and distance-to-centroid features yields both better classification and richer attention, suggesting that feature choice and explainability interact, not just accuracy.
  • CGAT's classification head can be swapped for a regression head, giving the same attention-map explanation mechanism for continuous tasks such as age estimation from other 3D shapes.
  • The 0.25 mean absolute error, lower than GAT's 0.34, matters for an ordinal staging task: errors are not just fewer but smaller in stage distance.

Reading between the lines

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

  • Editorial extension: the paper's central interpretability premise—that CLS attention weights equal decision contribution—can be directly tested by perturbation. If removing or deforming the highest-attention regions (root apices, furcation) does not change predictions more than removing low-attention regions, the maps would not be faithful explanations even though they look anatomically plausible.
  • Editorial extension: attention rollout assumes residual connections act as skip paths; in CGAT the residual structure differs from the transformers where rollout was introduced, so the composite weights may mix layers differently. Comparing rollout with gradient-based attribution on the same CGAT would clarify whether the two agree.
  • Editorial extension: the same CLS-attention mechanism could be applied to other high-stakes 3D graph tasks—such as bone age, tumor grading, or anatomical landmarking—where an expert-verifiable map of decision support might be as important as accuracy.
  • Editorial extension: because the dataset is small (528 teeth, 67% of samples in one stage) and class imbalance is extreme, the 0.76 F1 is not the main evidence; the paper's real test is whether the anatomical focus pattern generalizes to external or longitudinal data.
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 / 3 minor

Summary. The paper introduces the Class Node Graph Attention Network (CGAT), a GNN architecture that appends a virtual CLS node to graph-encoded 3D tooth meshes and uses GATv2 attention with attention rollout to generate per-node explanation maps. The method is evaluated on Demirjian dental stage classification (stages D–H) from CBCT-derived third molar meshes, using mean curvature and distance-to-centroid as node features. The authors report weighted F1 scores up to 0.76 and claim that the resulting attention maps align with human expert understanding, shifting focus from crown to furcation to roots and apices across stages. They compare against PointNet++, MeshCNN, GCN, and GAT baselines, and conclude that CGAT provides competitive accuracy combined with interpretable attention maps.

Significance. If the central explainability claim is valid, CGAT would be a useful step toward interpretable 3D shape classification in high-stakes settings, and the dental application is appropriate and societally relevant. The architecture is clearly described, and the paper systematically varies model depth, node features, and CLS edge direction. However, the validation of the attention maps as explanations is entirely qualitative, the classification evaluation omits standard deviations and does not use a patient-level split, and the faithfulness of attention rollout is not demonstrated. These gaps currently prevent the paper from supporting its main claim that the attention maps reflect the model's decision basis and align with human understanding.

major comments (4)
  1. [§2.2, Eq. (6) and Eq. (4)] The attention rollout recursion in Eq. (6) assumes information propagates approximately linearly through the network. However, the CGAT block in Eq. (4) contains GELU nonlinearities, LayerNorm, residual connections, and max-pooling over K=8 attention heads. Max-pooling is not a linear combination of heads, so there is no single attention matrix whose repeated product represents the flow of information to the CLS node. The paper does not provide any quantitative evidence that the rollout scores correspond to actual contribution to the classification decision. I would expect at least a comparison with gradient-based or perturbation-based attribution, or an ablation showing that removing high-attention nodes changes predictions more than removing low-attention nodes.
  2. [§2.2 (Model Training and Evaluation) and §3.2] The paper states that 'the conformity of these attention maps to human understanding is of higher priority than the classification metrics,' but no human evaluation is reported. The qualitative conclusions in §3.2 are based on the authors' visual inspection of a few sample teeth, and the statement that the maps 'align with human understanding' is therefore unsupported. The paper should either include a blinded expert rating study on the attention maps or provide quantitative alignment metrics (e.g., overlap with expert-annotated anatomical regions, or faithfulness metrics such as insertion/deletion). Without such validation, the central contribution of the paper is not established.
  3. [§2.1 (Data & Preprocessing)] The train/validation/test split is performed at the level of individual teeth, not patients. With 138 patients and 528 teeth (about 4 teeth per patient), teeth from the same patient almost certainly appear in both training and test sets, which can inflate performance and does not reflect the intended forensic use case of generalizing to new individuals. The experiments should be repeated with a patient-level split (e.g., a group split), and the baseline comparisons in Table 2 should be reported under the same split. Without this, the reported F1 scores may be optimistically biased.
  4. [Table 1 and Table 2] The paper reports mean weighted F1 scores over 10 runs but explicitly omits standard deviations 'for the sake of brevity.' This makes it impossible to assess whether differences such as 0.76 vs 0.73 are significant. Similarly, Table 2 reports single numbers for baselines and CGAT without confidence intervals or repeated-run variance. Please report standard deviations and, if claiming superiority, provide appropriate statistical tests. Additionally, specify the exact hyperparameters and node features used for the baseline models (GCN, GAT, MeshCNN, PointNet++); if baselines use different input representations or features, the comparison is not apples-to-apples.
minor comments (3)
  1. [Throughout] The paper contains several typos and formatting issues: 'T able 1' at the top of the table, 'N.Layers' as a column header, and the notation '1:15CGAT↔curv' is difficult to parse. The axes in Figure 6 are not labeled; Figures 7 and 8 would benefit from a color scale legend. Please proofread for consistency.
  2. [§3.3 and Conclusion] The claim of 'state-of-the-art performance' is too strong given the modest 0.76 weighted F1 on a small, imbalanced dataset and the absence of a patient-level evaluation. Similarly, the claim to be 'the first study to propose an architecture capable of generating attention-based explanations for its predictions on 3D shapes' needs a more careful literature check; at minimum, the statement should be softened or contextualized.
  3. [§2.2 (Attention Map Generation)] The description of attention rollout in Eq. (6) refers to a matrix A(l) of attention scores, but the preceding equations describe scalar attention coefficients per edge. Please clarify how the sparse attention matrix is constructed for the CLS node and how the max-pooling over heads is incorporated. The distinction between the attention to the CLS node and the attention between input nodes is also not fully clear in the rollout computation.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: attention maps are post-hoc visualizations of learned weights, classification accuracy is a genuine held-out result, and the only self-citations are not load-bearing.

full rationale

CGAT's classification performance is a genuine held-out result: models are trained with cross-entropy on 80/5/15 splits and compared against independent baselines (Table 2). No fitted parameter is renamed as a prediction; the attention maps are post-hoc visualizations of learned attention weights (Eq. 6), and the claim that they 'align with human understanding' is an empirical, qualitative claim supported by the authors' visual inspection of Figs. 7-8, not by a derivation that presupposes the conclusion. The derivation chain is self-contained: the CLS node is appended (Algorithm 1), attention weights are learned (Eq. 4), rollout composes them with the identity (Eq. 6), and predictions come from an MLP on the CLS embedding (Eq. 5). The architecture borrows known ingredients (GATv2, CLS/master node, attention rollout) with explicit citations to external work (Abnar & Zuidema, Dosovitskiy, Brody et al., Gilmer et al.); none of these is a self-citation chain that forces the reported result. The only self-citations ([25] Matthijs et al., [13] Franco et al., [40] Vranckx et al.) support a literature-gap claim, a staging-method citation, and dataset provenance; none is load-bearing for the claimed contribution. Weaknesses noted by the skeptic remain validity/design concerns, not circular reductions: attention rollout's faithfulness under max-pooling and GELU nonlinearities (Eq. 4 vs Eq. 6) is an unvalidated assumption about the explanation's correctness, and the qualitative evaluation of the maps is author-judged and not externally validated. But the paper never defines its explanation target in terms of the maps, and no fitted input later reappears as the 'predicted' explanation. Thus the central derivation is not circular; score 2 reflects only a minor, non-load-bearing self-citation and a subjective evaluation design.

Assumptions & free parameters 6 free parameters · 5 assumptions · 1 invented entities

The model rests on standard deep learning training (fitted weights are not listed; they are not free parameters in the scientific sense). The free parameters listed are the hyperparameters and preprocessing choices the authors selected by hand. The most important axioms are the sufficiency of the two hand-crafted node features and the faithfulness of attention as an explanation, the latter being the paper's load-bearing but unproven premise.

free parameters (6)
  • Number of CGAT blocks L = varied 1-15, optimal 6-12
    Model depth is a hyperparameter swept by the authors; chosen based on F1 and attention map quality.
  • Attention heads K = 8
    Set to 8 for all experiments, chosen by the authors without ablation.
  • Learning rate = 0.001
    Selected for Adam optimizer, standard choice.
  • Dropout probability = 0.3
    Applied to CLS embedding, chosen by hand.
  • Attention clipping range = [min(A), mean(A)+min(A)]
    Visualization-specific clipping to improve readability; affects the appearance of attention maps in figures.
  • Mesh decimation target = 751 nodes
    Meshes decimated to ~751 nodes; a preprocessing choice that could affect results.
assumptions (5)
  • domain assumption The undirected graph derived from the tooth mesh captures the shape information necessary for stage classification.
    Section 2.1 defines all meshes as undirected graphs; if topological detail is lost in decimation, the model's input is insufficient.
  • domain assumption Mean curvature and distance to centroid are sufficient and informative shape descriptors for dental stage.
    Section 2.3 selects these two features; the model's attention and classification depend entirely on them.
  • standard math Attention rollout composes per-layer attention weights into a meaningful measure of influence.
    Equation 6, following Abnar and Zuidema; assumes residual connections are modeled by adding identity.
  • domain assumption The Demirjian stage labels assigned by the two dentists are correct ground truth.
    Section 2.1, labels come from two experts with a tie-breaker; no inter-observer agreement statistics reported.
  • ad hoc to paper GATv2 attention weights can be interpreted as explanations of the model's decision.
    This is the load-bearing assumption for the explainability claim; not proven in the paper and contradicted by some literature.
invented entities (1)
  • CLS node (virtual node)
    purpose: Global information aggregator that collects attention from all mesh nodes and provides the classification embedding.
    The CLS node is a model component introduced in this paper (though similar to BERT's [CLS] and Graphormer's master node). It has no existence outside the model, so no independent falsifiable handle.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Attention Maps in 3D Shape Classification for Dental Stage Estimation with Class Node Graph Attention Networks." pith.science (2026). https://pith.science/paper/YJSQFETS

@misc{pith2026250907581,
  author       = {Pith},
  title        = {Pith review of: Attention Maps in 3D Shape Classification for Dental Stage Estimation with Class Node Graph Attention Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YJSQFETS}},
  note         = {Machine review of arXiv:2509.07581}
}
read the original abstract

Deep learning offers a promising avenue for automating many recognition tasks in fields such as medicine and forensics. However, the black-box nature of these models hinders their adoption in high-stakes applications where trust and accountability are required. For 3D shape recognition tasks in particular, this paper introduces the Class Node Graph Attention Network (CGAT) architecture to address this need. Applied to 3D meshes of third molars derived from CBCT images, for Demirjian stage allocation, CGAT utilizes graph attention convolutions and an inherent attention mechanism, visualized via attention rollout, to explain its decision-making process. We evaluated the local mean curvature and distance to centroid node features, both individually and in combination, as well as model depth, finding that models incorporating directed edges to a global CLS node produced more intuitive attention maps, while also yielding desirable classification performance. We analyzed the attention-based explanations of the models, and their predictive performances to propose optimal settings for the CGAT. The combination of local mean curvature and distance to centroid as node features yielded a slight performance increase with 0.76 weighted F1 score, and more comprehensive attention visualizations. The CGAT architecture's ability to generate human-understandable attention maps can enhance trust and facilitate expert validation of model decisions. While demonstrated on dental data, CGAT is broadly applicable to graph-based classification and regression tasks, promoting wider adoption of transparent and competitive deep learning models in high-stakes environments.

Figures

Figures reproduced from arXiv: 2509.07581 by the authors.

Figure 1
Figure 1. A visualization of the criteria for stage allocation of stages D to H according to the Demirjian assessment system [8]. The staging criteria in this range depend mostly on root development. As the development progresses, the roots start to protrude from the crown and get longer, finally developing closed apices. The dataset of focus for this study consists of triangular meshes of the four human third molars, corresp… view at source ↗
Figure 2
Figure 2. Distribution of samples over the Demirjian developmental stages. Note that the stages contain high class imbalance, with the lowest stage making up approximately 3% of all data, and the stage containing most samples, G, accounting for 67% of the dataset. five Belgian centers by Vranckx et al. in the form of CBCT images of prospective patients for third molar surgery [40]. Additional sample selection criteria were th… view at source ↗
Figure 3
Figure 3. Visualizations of sample meshes from each Demirjian stage in full resolution (a), and their decimated counterparts (b). The decimation allows for the elimination of the negative effects of the substantial variation in the number of nodes and edges, and makes model training and computational analyses feasible. It can be seen that the original outputs of the Relu® Creator software are overly dense, and the decimation … view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Overview of the CGAT architecture. The overall architecture consists of three sections: (1) The input section is responsible for connecting the virtual CLS node to all nodes in the input graph, and applying a linear projection of all node features. (2) The graph convol…
Figure 5
Figure 5. Figure 5: Distributions of the normalized mean curvature feature for all nodes, separated by Demirjian stage. As dental development progresses from stage D to H, the distribu￾tion skews, showing an increase in nodes with high positive curvature. This corresponds to the formation…
Figure 6
Figure 6. Figure 6: Plots of the weighted F1 scores of the models with the number of blocks rang￾ing from 1 to 15, trained using the mean curvature, the distance to centroid and both features combined. Overall, the CGAT model performance does not deteriorate signifi￾cantly in the given ra…
Figure 7
Figure 7. Figure 7: The attention maps from the 1,7,14CGAT{↔,→} family of the tested CGAT models, visualized on a sample tooth mesh of stage H. Across both undirected (↔) and directed (→) CLS edges, 1-block models exhibit noisy and mosaic attention, which becomes smoother and more focused…
Figure 8
Figure 8. Figure 8 [PITH_FULL_IMAGE:figures/full_fig_p020_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 31 canonical work pages

  1. [1]

    arXiv preprint arXiv:2005.00928 (2020)

    Abnar, S., Zuidema, W.: Quantifying attention flow in transformers. arXiv preprint arXiv:2005.00928 (2020)

    arXiv 2005

  2. [2]

    Computerized Medical Imaging and Graphics 91, 101939 (2021)

    Besson, P., Parrish, T., Katsaggelos, A.K., Bandt, S.K.: Geometric deep learning on brain shape predicts sex and age. Computerized Medical Imaging and Graphics 91, 101939 (2021)

    work page 2021

  3. [3]

    Brody, S., Alon, U., Yahav, E.: How attentive are graph attention networks? arXiv preprint arXiv:2105.14491 (2021)

    arXiv 2021

  4. [4]

    Dentomaxillofacial Radiology43(1), 20130291 (2014)

    Brown, J., Jacobs, R., Levring Jäghagen, E., Lindh, C., Baksi, G., Schulze, D., Schulze, R.: Basic training requirements for the use of dental cbct by dentists: a position paper prepared by the european academy of dentomaxillofacial radiology. Dentomaxillofacial Radiology43(1), 20130291 (2014)

    work page 2014

  5. [5]

    IEEE Access8, 35929–35949 (2020)

    Cao, W., Yan, Z., He, Z., He, Z.: A comprehensive survey on geometric deep learning. IEEE Access8, 35929–35949 (2020)

    work page 2020

  6. [6]

    IEEE transactions on image processing29, 3183–3198 (2019)

    Chen,S.,Duan,C.,Yang,Y.,Li,D.,Feng,C.,Tian,D.:Deepunsupervisedlearning of 3d point clouds via graph topology inference and filtering. IEEE transactions on image processing29, 3183–3198 (2019)

    work page 2019

  7. [7]

    In: Seminars in Musculoskeletal Radiology

    De Tobel, J., Ottow, C., Widek, T., Klasinc, I., Mörnstad, H., Thevissen, P.W., Verstraete, K.L.: Dental and skeletal imaging in forensic age estimation: disparities in current approaches and the continuing search for optimization. In: Seminars in Musculoskeletal Radiology. vol. 24, pp. 510–522. Thieme Medical Publishers (2020)

    work page 2020

  8. [8]

    Human biology pp

    Demirjian, A., Goldstein, H., Tanner, J.M.: A new system of dental age assessment. Human biology pp. 211–227 (1973)

    work page 1973

Show all 43 references
  1. [9]

    In: Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers)

    Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidi- rectional transformers for language understanding. In: Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologi...

    2019

  2. [10]

    IEEE Transactions on Tech- nology and Society4(1), 68–75 (2023)

    Dhar, T., Dey, N., Borra, S., Sherratt, R.S.: Challenges of deep learning in medical image analysis—improving explainability and trust. IEEE Transactions on Tech- nology and Society4(1), 68–75 (2023)

    2023

  3. [11]

    arXiv preprint arXiv:2010.11929 (2020) 24 Buyukcakir et al

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) 24 Buyukcakir et al

    2010 arXiv

  4. [12]

    Pattern Recog- nition Letters (2025)

    Facchi, G.M., Grossi, G., D’Amelio, A., Agnelli, F., Sforza, C., Tartaglia, G.M., Lanzarotti, R.: Graph neural networks for 3d facial morphology: Assessing the ef- fectiveness of anthropometric and automated landmark detection. Pattern Recog- nition Letters (2025)

    2025

  5. [13]

    International journal of legal medicine 134(1), 347–353 (2020)

    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. International journal of legal medicine 134(1), 347–353 (2020)

    2020

  6. [14]

    IEEE access 8, 57566–57593 (2020)

    Gezawa, A.S., Zhang, Y., Wang, Q., Yunqi, L.: A review on deep learning ap- proaches for 3d data representations in retrieval and classifications. IEEE access 8, 57566–57593 (2020)

    2020

  7. [15]

    In: International conference on machine learning

    Gilmer, J., Schoenholz, S.S., Riley, P.F., Vinyals, O., Dahl, G.E.: Neural message passing for quantum chemistry. In: International conference on machine learning. pp. 1263–1272. PMLR (2017)

    2017

  8. [16]

    In: Teh, Y.W., Titterington, M

    Glorot, X., Bengio, Y.: Understanding the difficulty of training deep feedforward neural networks. In: Teh, Y.W., Titterington, M. (eds.) Proceedings of the Thir- teenth International Conference on Artificial Intelligence and Statistics. Proceed- ings of Machine Learning Resea...

    2010

  9. [17]

    In: Proceedings of the IEEE/CVF international conference on computer vision workshops

    Gong, S., Chen, L., Bronstein, M., Zafeiriou, S.: Spiralnet++: A fast and highly ef- ficient mesh convolution operator. In: Proceedings of the IEEE/CVF international conference on computer vision workshops. pp. 0–0 (2019)

    2019

  10. [18]

    Advances in neural information processing systems30(2017)

    Hamilton, W., Ying, Z., Leskovec, J.: Inductive representation learning on large graphs. Advances in neural information processing systems30(2017)

    2017

  11. [19]

    ACM Transactions on Graphics (ToG)38(4), 1–12 (2019)

    Hanocka,R.,Hertz,A.,Fish,N.,Giryes,R.,Fleishman,S.,Cohen-Or,D.:Meshcnn: a network with an edge. ACM Transactions on Graphics (ToG)38(4), 1–12 (2019)

    2019

  12. [20]

    Dental Anthropology Journal18(2), 43–49 (2005)

    Harris, E.F., et al.: Tooth-coding systems in the clinical dental setting. Dental Anthropology Journal18(2), 43–49 (2005)

    2005

  13. [21]

    Information Fusion108, 102417 (2024)

    Hassanin, M., Anwar, S., Radwan, I., Khan, F.S., Mian, A.: Visual attention meth- ods in deep learning: An in-depth survey. Information Fusion108, 102417 (2024)

    2024

  14. [22]

    arXiv preprint arXiv:1609.02907 (2016)

    Kipf, T.N., Welling, M.: Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907 (2016)

    2016 arXiv

  15. [23]

    In: Chinese Conference on Biometric Recognition

    Liu, X., Yuan, L., Jiang, C., JiannanYu, Li, Y.: Human identification using tooth based on pointnet++. In: Chinese Conference on Biometric Recognition. pp. 129–

  16. [24]

    Ieee Access10, 23450–23462 (2022)

    Mahdi, S.S., Matthews, H., Nauwelaers, N., Vanneste, M., Gong, S., Bouritsas, G., Baynam, G.S., Hammond, P., Spritz, R., Klein, O.D., et al.: Multi-scale part-based syndrome classification of 3d facial images. Ieee Access10, 23450–23462 (2022)

    2022

  17. [25]

    International Journal of Legal Medicine138(6), 2469–2479 (2024)

    Matthijs, L., Delande, L., De Tobel, J., Büyükçakir, B., Claes, P., Vandermeulen, D., Thevissen, P.: Artificial intelligence and dental age estimation: development and validation of an automated stage allocation technique on all mandibular tooth types in panoramic radiographs....

    2024

  18. [26]

    Artificial Intelligence Review56(8), 8647– 8695 (2023)

    Minaee, S., Abdolrashidi, A., Su, H., Bennamoun, M., Zhang, D.: Biometrics recog- nition using deep learning: A survey. Artificial Intelligence Review56(8), 8647– 8695 (2023)

    2023

  19. [27]

    Applied Sciences14(8), 3424 (2024)

    Mohammadi, S., Allali, M.: Advancing brain tumor segmentation with spectral– spatial graph neural networks. Applied Sciences14(8), 3424 (2024)

    2024

  20. [28]

    Select Committee on Artificial In- telligence: The National Artificial Intelligence Research and Development Strategic Class Node Graph Attention Networks 25 Plan: 2023 Update

    National Science and Technology Council (US). Select Committee on Artificial In- telligence: The National Artificial Intelligence Research and Development Strategic Class Node Graph Attention Networks 25 Plan: 2023 Update. National Science and Technology Council (US), Select C...

    2023

  21. [29]

    Piccialli, F., Di Somma, V., Giampaolo, F., Cuomo, S., Fortino, G.: A survey on deep learning in medicine: Why, how and when? Information Fusion66, 111–137 (2021)

    2021

  22. [30]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 652–660 (2017)

    2017

  23. [31]

    Advances in neural information processing systems30(2017)

    Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learn- ing on point sets in a metric space. Advances in neural information processing systems30(2017)

    2017

  24. [32]

    Forensic Science International346, 111654 (2023)

    Rahim, A.A., Davies, J., Liversidge, H.: Reliability and limitations of permanent tooth staging techniques. Forensic Science International346, 111654 (2023)

    2023

  25. [33]

    https://www.relu.eu/creator (2022–2025)

    Relu BV: Relu Creator. https://www.relu.eu/creator (2022–2025)

    2022

  26. [34]

    arXiv preprint arXiv:2303.10993 (2023)

    Rusch, T.K., Bronstein, M.M., Mishra, S.: A survey on oversmoothing in graph neural networks. arXiv preprint arXiv:2303.10993 (2023)

    2023 arXiv

  27. [35]

    In: Panel for the Future of Science and Technology (2020)

    Sartor, G., Lagioia, F.: The impact of the general data protection regulation (gdpr) on artificial intelligence. In: Panel for the Future of Science and Technology (2020)

    2020

  28. [36]

    In: Geometric Deep Learning in Medical Image Analysis

    Shehata, N., Bain, W., Glocker, B.: A comparative study of graph neural networks for shape classification in neuroimaging. In: Geometric Deep Learning in Medical Image Analysis. pp. 160–171. PMLR (2022)

    2022

  29. [37]

    Advances in neural information pro- cessing systems30(2017)

    Vaswani,A.,Shazeer,N.,Parmar,N.,Uszkoreit,J.,Jones,L.,Gomez,A.N.,Kaiser, Ł., Polosukhin, I.: Attention is all you need. Advances in neural information pro- cessing systems30(2017)

    2017

  30. [38]

    stat1050(20), 10–48550 (2017)

    Velickovic, P., Cucurull, G., Casanova, A., Romero, A., Lio, P., Bengio, Y., et al.: Graph attention networks. stat1050(20), 10–48550 (2017)

    2017

  31. [39]

    Vosylius, V., Wang, A., Waters, C., Zakharov, A., Ward, F., Le Folgoc, L., Cupitt, J., Makropoulos, A., Schuh, A., Rueckert, D., et al.: Geometric deep learning for post-menstrual age prediction based on the neonatal white matter cortical surface. In: Uncertainty for Safe Util...

    2020

  32. [40]

    symptomatic third molar removal: effects on patient postoperative morbidity

    Vranckx, M., Fieuws, S., Jacobs, R., Politis, C.: Prophylactic vs. symptomatic third molar removal: effects on patient postoperative morbidity. Journal of Evidence Based Dental Practice21(3), 101582 (2021)

    2021

  33. [41]

    Science China Life Sciences67(7), 1489–1501 (2024)

    Yang, X., Li, R., Yang, X., Zhou, Y., Liu, Y., Han, J.D.J.: Coordinate-wise mono- tonic transformations enable privacy-preserving age estimation with 3d face point cloud. Science China Life Sciences67(7), 1489–1501 (2024)

    2024

  34. [42]

    In: 2024 IEEE International Symposium on Biomedical Imaging (ISBI)

    Yang, Z., Su, Y., Rouhi, R., Farazi, M., Zhu, W., Chen, Y., Reiman, E.M., Caselli, R.J., Chen, K., Wang, Y., et al.: Pre-training graph attention convolution for brain structural imaging biomarker analysis and its application to alzheimer’s disease pathology identification. In...

    2024

  35. [43]

    Ying, C., Cai, T., Luo, S., Zheng, S., Ke, G., He, D., Shen, Y., Liu, T.Y.: Do transformers really perform badly for graph representation? Advances in neural information processing systems34, 28877–28888 (2021)

    2021

Pith tools

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