Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

Structured Spectral Graph Learning for Anomaly Classification in 3D Chest CT Scans

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read CT-Graph models a 3D chest CT scan as a graph of slice triplets and uses spectral convolution to outperform prior methods on multi-label anomaly classification.

desk verdict Incremental but honest architecture paper; the main F1 gain on CT-RATE holds, but the cross-dataset and 'consistently outperforms' claims are overreached. read the letter →

arxiv 2508.01045 v1 pith:2JQ6WVIE submitted 2025-08-01 cs.CV cs.LG

Theo Di Piazza , Carole Lazarus , Olivier Nempont , Loic Boussel This is my paper
classification cs.CVcs.LG
keywords 3DchestCTmulti-labelanomalyclassificationgraphneuralnetworkChebyshevconvolutionspectralslicetripletsCT-RATERad-Chest
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 claims that a 3D chest CT volume can be cast as a structured graph, with one node per triplet of adjacent axial slices and edges weighted by physical z-axis distance, and that propagating features through Chebyshev (spectral) graph convolutions improves multi-label anomaly classification. The proposed CT-Graph obtains F1-Score 54.59 and AUROC 82.44 on the CT-RATE test set, ahead of all compared baselines, and F1-Score 49.52 and AUROC 72.18 on the Rad-ChestCT dataset, supporting the paper's cross-dataset generalization claim. Ablations identify the spectral convolution, the neighborhood-constrained graph, and the inverse-distance edge weighting as the components responsible for the gains, along with robustness to z-axis translation.

What carries the argument

The load-bearing machinery is CT-Graph's graph construction and its spectral message passing. Each scan is partitioned into 80 non-overlapping triplets of axial slices; a pretrained ResNet with global average pooling turns each triplet into a 512-dimensional node feature. Edges connect nodes at most $q$ positions apart, weighted by $w_{i,j} = 1 + \frac{1}{1 + 3|i-j|s_z}$, where $s_z$ is the z-axis spacing in decimeters, so the graph encodes physical proximity with no extra learned parameters. Three Chebyshev convolution layers, built from polynomial approximations of the normalized graph Laplacian, propagate features across this graph in the spectral domain; the node outputs are summed and fed to an MLP classification head. The spectral filter is what carries the argument: unlike spatial graph convolution it does not rely on a fixed local neighborhood, so long-range axial dependencies can inform each node without assuming the same anatomy appears at the same slice index in every patient.

What would settle it

Have two radiologists independently re-label a random sample of Rad-ChestCT volumes using the CT-RATE label definitions and measure per-label agreement with the original Rad-ChestCT labels; if several of the 16 shared anomalies disagree, the reported cross-dataset F1 and AUROC compare inconsistent labels, and the generalization claim would need to be re-tested on a remapped subset.

Watch

Extended reading notes

Core claim

The central discovery is that spectral-domain graph convolution over slice-triplet nodes is a viable alternative to 3D convolutions and transformers for chest CT anomaly classification: it matches or beats strong 2.5D baselines while remaining robust to patient-positioning variation along the z-axis. Concretely, CT-Graph reports F1-Score 54.59 with AUROC 82.44 on CT-RATE (versus 53.97 and 81.80 for CT-Scroll) and F1-Score 49.52 with AUROC 72.18 on Rad-ChestCT (versus 49.28 and 71.21 for the next best baseline, CNN3D). The authors also demonstrate with synthetic translations that the spectral model, unlike standard spatial convolution, degrades gracefully when the volume is shifted along the z-axis, which they attribute to the Laplacian-based filtering's independence from fixed anatomical alignment.

Load-bearing premise

The cross-dataset generalization claim assumes the 16 Rad-ChestCT anomaly labels match the CT-RATE taxonomy in meaning and annotation criteria, yet the paper gives no explicit label mapping or harmonization procedure for the two datasets.

Editorial extensions

If this is right

  • On CT-RATE, CT-Graph reports the best F1-Score (54.59), AUROC (82.44), and Recall (68.77) among all compared methods, with the F1 gain over each baseline significant at p<0.01.
  • On Rad-ChestCT, it again takes the top F1-Score (49.52) and AUROC (72.18), the basis for the paper's claim that the model generalizes across datasets without same-domain pretraining.
  • The ablation shows spectral (Chebyshev) convolution outperforms spatial GraphConv, so the spectral domain contributes beyond graph connectivity alone.
  • Constraining connectivity to a neighborhood (q=16) and weighting edges by inverse physical distance improve results over fully connected, unweighted graphs.
  • The model holds its F1 score under artificial z-axis volume shifts, indicating reliability when patient positioning in the scanner varies.

Reading between the lines

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

  • A natural next experiment, not run in the paper, is to stress the z-axis robustness claim with larger translations or non-uniform resampling, to find the displacement envelope where spectral convolution's advantage over spatial GNNs disappears.
  • If the label-compatibility assumption holds, the same parameter-free edge-weighting scheme could transfer to other volumetric modalities with variable coverage, such as MRI or cone-beam CT.
  • The graph abstraction suggests an untested hybrid: inject anatomical segmentation maps as node or edge features, connecting nodes that are physically adjacent in 3D even when far apart along the z-axis.
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 / 6 minor

Summary. The paper proposes CT-Graph, a graph-based architecture for multi-label anomaly classification in 3D chest CT volumes. Each scan is represented as a graph whose nodes are triplets of adjacent axial slices encoded by an ImageNet-pretrained ResNet; edges connect nearby slice triplets with weights that decay with physical z-axis distance. A three-layer Chebyshev (spectral) graph convolution network aggregates node features before a lightweight MLP head performs multi-label classification. Experiments on the CT-RATE test set report an F1 score of 54.59, AUROC of 82.44, and recall of 68.77 over five runs, with a paired t-test p<0.01 for F1 against baselines. The paper also reports results on Rad-ChestCT (F1 49.52, AUROC 72.18) as evidence of cross-dataset generalization, ablation studies on graph connectivity, neighbourhood size, and edge weighting functions, and a z-axis translation robustness experiment.

Significance. If the CT-RATE results hold, CT-Graph is a credible alternative to 3D CNNs and attention-based 2.5D methods for volumetric anomaly classification, with a modest but consistent F1/AUROC improvement over the strongest prior baseline (CT-Scroll) and a lightweight inference cost of about 70 ms. The paper has notable strengths: evaluation on two public datasets, five independent runs with a paired significance test, and systematic ablations of graph construction and GNN module choices. However, the significance is tempered by three issues: the claimed cross-dataset generalization relies on an unspecified label harmonization between CT-RATE and Rad-ChestCT; a blanket claim of consistent superiority across all metrics is contradicted by the paper's own Table 1; and the edge-weighting design is selected by inspecting test-set F1, so it is not a parameter-free contribution as stated. The core CT-RATE result may survive these concerns, but the headline claims need to be substantially re-framed.

major comments (4)
  1. [4.1] The cross-dataset evaluation is not fully specified. Section 4.1 states that Rad-ChestCT is used 'focusing on the 16 anomalies shared with CT-RATE', but no label mapping, harmonization procedure, or manual verification is provided. CT-RATE labels are derived from radiology reports, while Rad-ChestCT uses a different annotation pipeline; the same named anomaly may have different diagnostic criteria. Because the abstract's claim of 'strong cross-dataset generalization' rests on the Rad-ChestCT results in Table 1, the authors should provide the exact label mapping, any verification by a radiologist or automated agreement check, and a sensitivity analysis to alternative mapping choices. Without this, the roughly one-point F1/AUROC gains on Rad-ChestCT are not interpretable as evidence of generalization.
  2. [4.3] The statement in Section 4.3 that 'CT-Graph consistently outperforms all baselines across AUROC, F1-Score and Recall' is contradicted by Table 1. On Rad-ChestCT, CNN3D has higher recall (70.47 vs. 69.30) and CT-Scroll has higher accuracy (63.02 vs. 62.60); on CT-RATE, CT-Scroll also has higher accuracy (79.49 vs. 78.66). The claim should be revised to name the specific metrics and datasets on which CT-Graph is consistently best (F1 and AUROC), and the statistical significance statement should be extended to those metrics or restricted accordingly.
  3. [4.4, Figure 4] The edge-weighting function is selected using the test set. Figure 4 reports results 'on the CT-RATE test set' and the inverse function with z-axis spacing in decimeters is chosen because it yields the best F1 score. This means the specific form of Eq. (5), including the constant 3 and the choice of units, is a test-set-tuned hyperparameter rather than a parameter-free design as claimed in the Introduction. The authors should either move this model selection to the validation set, pre-register the weighting scheme, or report sensitivity on a held-out split; otherwise the claimed advantage of the weighting function may reflect overfitting to the test set.
  4. [4.3, Figure 3b] The z-axis translation robustness experiment is insufficiently described for reproduction or assessment. The text and Figure 3b do not state the shift magnitudes used, the padding strategy, which method is identified as 'standard convolution', or whether the reported F1 values are averaged over multiple shift amounts. Since robustness to z-axis translation is a stated contribution of the paper, the authors should provide a precise protocol, including shift values, padding type, and per-shift results for all compared methods.
minor comments (6)
  1. [4.3] The reported gain of '+∆1.15%' over CT-Scroll is relative (0.62 points absolute F1, from 53.97 to 54.59); this should be stated unambiguously to avoid confusing readers.
  2. [3, Eq. (5)] The notation dist(i,j) is defined only through the subsequent formula; clarifying that dist(i,j)=3×|i−j|×sz, with sz in decimeters, would improve readability.
  3. [2.1] There are several typos, including 'primaliry' (should be 'primarily'), 'independant' in the Table 1 caption, and 'Reffering' in Section 4.3. A careful proofread is needed.
  4. [4.2] No hardware details are given for the reported inference time of approximately 70 ms; adding the GPU model would make the efficiency claim reproducible.
  5. [4.4] The neighbourhood size q=16 is used in Table 2, but the choice of q is not justified; since Table 3 shows only small differences across q values, a brief comment on sensitivity to q would be useful.
  6. [References] Reference [11] (CT-Scroll) is the authors' own prior work and is used as the main comparison baseline; the paper should explicitly disclose this relationship in the main text for transparency.

Circularity Check

2 steps flagged · score 6.0 of 10

The edge-weighting scheme and GNN module are selected by ranking F1 on the CT-RATE test set, so the reported superiority of CT-Graph is partly self-confirming rather than an independent prediction.

  1. fitted input called prediction [Section 4.4, Figure 4]
    "Among the evaluated edge weighting functions, the inverse function (see Eq. 5) with z-axis spacing measured in decimeters (dm) yields the best classification performance, as illustrated in Figure 4. ... Fig.4: Impact of the edge weighting functions, on the CT-RATE test set."

    The inverse-distance edge weighting and its chosen unit (dm) are selected because they maximize F1 on the CT-RATE test set. The same test set is then used in Table 1 to claim that CT-Graph 'consistently outperforms all baselines.' Thus the reported optimality of Eq. 5 is not an a priori design prediction; it is a restatement of the test-set ranking used to select the weighting. This is a fitted design choice presented as a validated contribution, making the final test-set comparison partially self-confirming.

  2. fitted input called prediction [Section 4.4, Table 2]
    "Comparison of representative GNNs. Table 2 highlights the performance gains achieved by incorporating Chebyshev Convolutions [10] in our GNN module. Compared to a direct neighborhood aggregation approach [27], ChebConv improves AUROC by +∆0.42% and F1-Score by +∆1.25% ... Table2: Comparison of graph connectivity schemes and GNN modules, evaluated on the CT-RATE test set."

    The choice of ChebConv over GraphConv and GATv2Conv is justified by its AUROC and F1 advantage on the CT-RATE test set, with the table caption explicitly stating that the comparison is evaluated on the CT-RATE test set. Since the final model incorporates this module chosen by inspecting the same test set, the conclusion that 'spectral-domain convolutions may enhance feature aggregation' is a description of the selection criterion rather than an independent confirmation. This is a milder form of circularity than equation-level reduction, but it still means the reported design superiority is partly determined by the evaluation set.

full rationale

The core architecture is not derived from the target results: CT-Graph is a graph model with spectral convolution, and its headline CT-RATE and Rad-ChestCT numbers are compared against external baselines, so the main benchmark comparison is not circular. However, two key design choices—the edge-weighting function with decimeter units and the Chebyshev convolution module—are selected by ranking performance on the CT-RATE test set, and the same test set is then used to claim state-of-the-art results. This is the 'fitted input called prediction' pattern: the optimality of those design elements is a restatement of the test-set ranking used to choose them, which partially inflates the reported advantage. The cross-dataset label-mapping concern is a validity issue rather than circularity, and the self-citation of CT-Scroll is not load-bearing. Overall, the central claim retains independent empirical content, but the specific 'best' design conclusions are self-confirming, giving a partial circularity score of 6.

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

No new physical entities, particles, forces, or conserved quantities are introduced; the graph nodes and weights are representational choices, not new entities in the world. The free parameters are hand-chosen hyperparameters that the reported performance depends on.

free parameters (5)
  • Z-spacing unit for edge weights = decimetre (dm)
    Eq. 5 uses sz in decimeters; Figure 4 shows this unit choice gives best test F1. The unit is a free scaling hyperparameter selected on the test set.
  • Edge weight constant (1+1/(1+3|i-j|sz)) = As in Eq. 5
    Hand-designed inverse-distance weighting; the functional form and the factor 3 are not derived from data.
  • Neighborhood size q = 16
    Table 3 compares q=4,16,80 and 16 is used for the main results; chosen post hoc on the test set.
  • Chebyshev filter size K = 3
    Fixed to 3 'to align with common practice [10]' (Section 3).
  • Number of GNN layers = 3
    Set to match CT-Scroll depth 'for fair comparison' (Section 3).
assumptions (4)
  • domain assumption CT volumes can be represented as a graph whose nodes are triplets of adjacent axial slices.
    Core modeling choice, stated in Section 3 (Graph Construction).
  • standard math Chebyshev polynomial approximation of graph filters is a valid and sufficient convolution operator for this graph.
    Borrowed from Defferrard et al. [10]; invoked in Eq. 6.
  • domain assumption ImageNet-pretrained ResNet features summarize each slice triplet well enough for anomaly classification.
    All methods use this feature extractor; no fine-tuning is described. Section 3, Eq. 1.
  • domain assumption Rad-ChestCT label definitions align with CT-RATE for the 16 shared anomalies.
    Section 4.1 states the shared anomalies but provides no harmonization details.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Structured Spectral Graph Learning for Anomaly Classification in 3D Chest CT Scans." pith.science (2026). https://pith.science/paper/2JQ6WVIE

@misc{pith2026250801045,
  author       = {Pith},
  title        = {Pith review of: Structured Spectral Graph Learning for Anomaly Classification in 3D Chest CT Scans},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2JQ6WVIE}},
  note         = {Machine review of arXiv:2508.01045}
}
read the original abstract

With the increasing number of CT scan examinations, there is a need for automated methods such as organ segmentation, anomaly detection and report generation to assist radiologists in managing their increasing workload. Multi-label classification of 3D CT scans remains a critical yet challenging task due to the complex spatial relationships within volumetric data and the variety of observed anomalies. Existing approaches based on 3D convolutional networks have limited abilities to model long-range dependencies while Vision Transformers suffer from high computational costs and often require extensive pre-training on large-scale datasets from the same domain to achieve competitive performance. In this work, we propose an alternative by introducing a new graph-based approach that models CT scans as structured graphs, leveraging axial slice triplets nodes processed through spectral domain convolution to enhance multi-label anomaly classification performance. Our method exhibits strong cross-dataset generalization, and competitive performance while achieving robustness to z-axis translation. An ablation study evaluates the contribution of each proposed component.

Figures

Figures reproduced from arXiv: 2508.01045 by the authors.

Figure 1
Figure 1. CT-Graph introduces a structured graph-based architecture, where triplet [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. GradCAM activation maps extracted from the 2D ResNet module. [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. (a) Per-anomaly F1-Score comparison for the 3 anomalies with highest [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Impact of the edge weighting functions, on the [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Structured Spectral Graph Representation Learning for Multi-label Abnormality Analysis from 3D CT Scans

    cs.CV 2025-10 conditional novelty 4.0 of 10

    A graph-of-slice-triplets encoder with spectral convolution outperforms 3D CNN/Transformer baselines on multi-label chest CT abnormality classification and transfers to report generation and abdominal CT.

Reference graph

Works this paper leans on

36 extracted references · 23 canonical work pages · cited by 1 Pith paper

  1. [1]

    Informatics in Medicine Unlocked26, 100723 (Jan 2021)

    Anaya-Isaza,A.,Mera-Jiménez,L.,Zequera-Diaz,M.:Anoverviewofdeeplearning in medical imaging. Informatics in Medicine Unlocked26, 100723 (Jan 2021)

    work page 2021

  2. [2]

    In: 2021 IEEE/CVF International Conference on Computer Vision (ICCV)

    Arnab, A., Dehghani, M., Heigold, G., Sun, C., Lucic, M., Schmid, C.: ViViT: A Video Vision Transformer. In: 2021 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 6816–6826. IEEE, Montreal, QC, Canada (Oct 2021)

    work page 2021

  3. [3]

    Medical Image Analysis91, 103000 (Jan 2024)

    Azad, R., Kazerouni, A., Heidari, M., Aghdam, E.K., Molaei, A., Jia, Y., Jose, A., Roy, R., Merhof, D.: Advances in medical image analysis with vision Transformers: A comprehensive review. Medical Image Analysis91, 103000 (Jan 2024)

    work page 2024

  4. [4]

    Bechler-Speicher, M., Globerson, A., Gilad-Bachrach, R.: The Intelligible and Ef- fective Graph Neural Additive Networks (Dec 2024), arXiv:2406.01317 [cs]

    work page arXiv 2024

  5. [5]

    In: Advances in Neural Information Processing Systems

    Belkin, M., Niyogi, P.: Laplacian Eigenmaps and Spectral Techniques for Em- bedding and Clustering. In: Advances in Neural Information Processing Systems. vol. 14. MIT Press (2001)

    work page 2001

  6. [6]

    Emergency Radiology13(1), 25–30 (Oct 2006)

    Broder, J., Warshauer, D.M.: Increasing utilization of computed tomography in the adult emergency department, 2000-2005. Emergency Radiology13(1), 25–30 (Oct 2006)

    work page 2000

  7. [7]

    Brody, S., Alon, U., Yahav, E.: How Attentive are Graph Attention Networks? (Jan 2022), arXiv:2105.14491 [cs]

    arXiv 2022

  8. [8]

    Bruna, J., Zaremba, W., Szlam, A., LeCun, Y.: Spectral Networks and Locally Connected Networks on Graphs (May 2014), arXiv:1312.6203 [cs]

    arXiv 2014

Show all 36 references
  1. [9]

    Chen, D.Z., Hu, R., Chen, X., Nießner, M., Chang, A.X.: UniT3D: A Uni- fied Transformer for 3D Dense Captioning and Visual Grounding (Dec 2022), arXiv:2212.00836 [cs]

    2022 arXiv

  2. [10]

    Defferrard, M., Bresson, X., Vandergheynst, P.: Convolutional Neural Networks on Graphs with Fast Localized Spectral Filtering (Feb 2017), arXiv:1606.09375 [cs]

    2017 arXiv

  3. [11]

    Di Piazza, T., Lazarus, C., Nempont, O., Boussel, L.: Imitating Radiological Scrolling: A Global-Local Attention Model for 3D Chest CT Volumes Multi-Label Anomaly Classification (Jan 2025)

    2025

  4. [12]

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale (Jun 2021), arXiv:2010.11929 [cs] ...

    2021 arXiv

  5. [13]

    Medical Image Analysis67, 101857 (Jan 2021)

    Draelos,R.L.,Dov,D.,Mazurowski,M.A.,Lo,J.Y.,Henao,R.,Rubin,G.D.,Carin, L.: Machine-learning-based multiple abnormality prediction with large-scale chest computed tomography volumes. Medical Image Analysis67, 101857 (Jan 2021)

    2021

  6. [14]

    Guo, Z., Zhao, W., Wang, S., Yu, L.: HIGT: Hierarchical Interaction Graph- Transformer for Whole Slide Image Analysis (Sep 2023), arXiv:2309.07400 [cs]

    2023 arXiv

  7. [15]

    Scientific Reports 14(1), 12567 (May 2024), publisher: Nature Publishing Group

    Halder, A., Gharami, S., Sadhu, P., Singh, P.K., Woźniak, M., Ijaz, M.F.: Imple- menting vision transformer for classifying 2D biomedical images. Scientific Reports 14(1), 12567 (May 2024), publisher: Nature Publishing Group

    2024

  8. [16]

    Hamamci, I.E., Er, S., Almas, F., Simsek, A.G., Esirgun, S.N., Dogan, I., Das- delen, M.F., Wittmann, B., Simsar, E., Simsar, M., Erdemir, E.B., Alanbay, A., Sekuboyina, A., Lafci, B., Ozdemir, M.K., Menze, B.: A foundation model utilizing chest CT volumes and radiology report...

    2024

  9. [17]

    Hamamci, I.E., Er, S., Menze, B.: CT2Rep: Automated Radiology Report Gener- ation for 3D Medical Imaging (Mar 2024), arXiv:2403.06801 [cs, eess]

    2024 arXiv

  10. [18]

    Hamamci, I.E., Er, S., Simsar, E., Sekuboyina, A., Prabhakar, C., Tezcan, A., Simsek, A.G., Esirgun, S.N., Almas, F., Doğan, I., Dasdelen, M.F., Reynaud, H., Pati, S., Bluethgen, C., Ozdemir, M.K., Menze, B.: GenerateCT: Text-Conditional Generation of 3D Chest CT Volumes (Nov ...

    2023 arXiv

  11. [19]

    He, K., Zhang, X., Ren, S., Sun, J.: Deep Residual Learning for Image Recognition (Dec 2015), arXiv:1512.03385 [cs]

    2015 arXiv

  12. [20]

    Computers in Biology and Medicine132, 104348 (May 2021)

    Ibrahim, D.M., Elshennawy, N.M., Sarhan, A.M.: Deep-chest: Multi-classification deep learning model for diagnosing COVID-19, pneumonia, and lung cancer chest diseases. Computers in Biology and Medicine132, 104348 (May 2021)

    2021

  13. [21]

    Heliyon10(6), e27398 (Mar 2024)

    Ilesanmi, A.E., Ilesanmi, T.O., Ajayi, B.O.: Reviewing 3D convolutional neural network approaches for medical image segmentation. Heliyon10(6), e27398 (Mar 2024)

    2024

  14. [22]

    Kiechle, J., Lang, D.M., Fischer, S.M., Felsner, L., Peeken, J.C., Schnabel, J.A.: Graph Neural Networks: A suitable Alternative to MLPs in Latent 3D Medical Image Classification? (Jul 2024), arXiv:2407.17219 [cs]

    2024 arXiv

  15. [23]

    Liu, F., Wu, X., Ge, S., Fan, W., Zou, Y.: Exploring and Distilling Posterior and Prior Knowledge for Radiology Report Generation (Jun 2021), arXiv:2106.06963 [cs]

    2021 arXiv

  16. [24]

    Liu, Z., Ning, J., Cao, Y., Wei, Y., Zhang, Z., Lin, S., Hu, H.: Video Swin Trans- former (Jun 2021), arXiv:2106.13230 [cs]

    2021 arXiv

  17. [25]

    Ma, J., Li, F., Wang, B.: U-Mamba: Enhancing Long-range Dependency for Biomedical Image Segmentation (Jan 2024), arXiv:2401.04722 [eess]

    2024 arXiv

  18. [26]

    Makarov, N., Narayanan, S., Antoniou, C.: Graph neural network surrogate for strategic transport planning (Aug 2024), arXiv:2408.07726 [cs]

    2024 arXiv

  19. [27]

    Morris, C., Ritzert, M., Fey, M., Hamilton, W.L., Lenssen, J.E., Rattan, G., Grohe, M.: Weisfeiler and Leman Go Neural: Higher-order Graph Neural Networks (Nov 2021), arXiv:1810.02244 [cs]

    2021 arXiv

  20. [28]

    Informatics in Medicine Unlocked47, 101504 (Jan 2024)

    Rayed, M.E., Islam, S.M.S., Niha, S.I., Jim, J.R., Kabir, M.M., Mridha, M.F.: Deep learning for medical image segmentation: State-of-the-art advancements and challenges. Informatics in Medicine Unlocked47, 101504 (Jan 2024)

    2024

  21. [29]

    Communications Materials3(1), 1–18 (Nov 2022), publisher: Nature Publishing Group Spectral Graph for 3D CT Scans 11

    Reiser, P., Neubert, M., Eberhard, A., Torresi, L., Zhou, C., Shao, C., Metni, H., van Hoesel, C., Schopmans, H., Sommer, T., Friederich, P.: Graph neural networks for materials science and chemistry. Communications Materials3(1), 1–18 (Nov 2022), publisher: Nature Publishing ...

    2022

  22. [30]

    Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., Berg, A.C., Fei-Fei, L.: ImageNet Large Scale Visual Recognition Challenge (Jan 2015), arXiv:1409.0575 [cs]

    2015 arXiv

  23. [31]

    Tucudean, G., Bucos, M., Dragulescu, B., Caleanu, C.D.: Natural language pro- cessing with transformers: a review. PeerJ. Computer Science10, e2222 (2024)

    2024

  24. [32]

    Vaswani,A.,Shazeer,N.,Parmar,N.,Uszkoreit,J.,Jones,L.,Gomez,A.N.,Kaiser, L., Polosukhin, I.: Attention Is All You Need (Aug 2023), arXiv:1706.03762 [cs]

    2023 arXiv

  25. [33]

    Current Opinion in Structural Biology79, 102538 (Apr 2023), arXiv:2301.08210 [cs]

    Veličković, P.: Everything is Connected: Graph Neural Networks. Current Opinion in Structural Biology79, 102538 (Apr 2023), arXiv:2301.08210 [cs]

    2023 arXiv

  26. [34]

    Yang, Y.Q., Guo, Y.X., Xiong, J.Y., Liu, Y., Pan, H., Wang, P.S., Tong, X., Guo, B.: Swin3D: A Pretrained Transformer Backbone for 3D Indoor Scene Understand- ing (Aug 2023), arXiv:2304.06906 [cs]

    2023 arXiv

  27. [35]

    Proceedings of SPIE–the International So- ciety for Optical Engineering12463, 124634J (Feb 2023)

    Yuan, Y., Tivnan, M., Gang, G.J., Stayman, J.W.: Deep Learning CT Image Restoration using System Blur Models. Proceedings of SPIE–the International So- ciety for Optical Engineering12463, 124634J (Feb 2023)

    2023

  28. [36]

    Zhang, Y., Huang, S.C., Zhou, Z., Lungren, M.P., Yeung, S.: Adapting Pre-trained Vision Transformers from 2D to 3D through Weight Inflation Improves Medical Image Segmentation (Feb 2023), arXiv:2302.04303 [cs]

    2023 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.