Pith. sign in

REVIEW 4 major objections 6 minor 31 references

For TCM tongue diagnosis, careful labels, restrained color augmentation, and more data beat fancy architectures—and 45 fine-grained labels collapse the model.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-31 16:23 UTC pith:CIPZ5H2H

load-bearing objection Solid small-N ablation cookbook with honest negatives; the big +20.6% scaling and 13→45 collapse claims are confounded and rest on noisy non-expert labels. the 4 major comments →

arxiv 2607.28148 v1 pith:CIPZ5H2H submitted 2026-07-30 cs.CV cs.LG

What Makes Deep Learning Work for Traditional Chinese Medicine Tongue Diagnosis? A Comprehensive Ablation Study

Longxia Gao , Linan Wang , Yuhe Han , Junze Geng , Meng Zhang , Hanqing Zhao This is my paper
classification cs.CV cs.LG
keywords tongue diagnosistraditional Chinese medicinedeep learningmulti-label classificationablation studyclass imbalanceConvNeXtmedical image analysis
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper asks what actually makes deep learning work for automated tongue diagnosis in traditional Chinese medicine, instead of assuming a bigger or more elaborate model is better. Across more than twenty controlled model versions under five-fold cross-validation, the authors show that a mid-sized ConvNeXt backbone, simple positive-weighted binary cross-entropy, very mild color jitter, and specialist classifiers for weak label groups outperform more aggressive losses, heavy augmentation, and complex training tricks. Scaling from under a thousand expert images to about eleven thousand samples with thirteen binary labels lifts weighted F1 by roughly twenty percent, while expanding to forty-five fine-grained labels crashes performance from about 0.78 to 0.22. The practical message is that data scale, label design, and domain-aware restraint matter more than architectural novelty for imbalanced multi-label medical images. A sympathetic reader cares because these choices are transferable design rules, not one-off model tuning.

Core claim

Under matched five-fold cross-validation, six design principles govern multi-label TCM tongue models: ConvNeXt-Tiny is the efficient backbone; BCE with positive-class weighting beats Asymmetric Loss by about 2.7 weighted-F1 points on sigmoid groups; restrained hue jitter near ±0.01 is required when color is diagnostic; replacing weak-group outputs with independent specialist classifiers gains about 2.1 points and beats probability averaging; scaling from 976 to 11,101 samples with 13 binary labels yields about +20.6% relative weighted-F1 (to 0.7761); and expanding from 13 to 45 label dimensions collapses weighted-F1 from about 0.78 to 0.22. The authors present these as generalizable to imbal

What carries the argument

A one-variable-at-a-time ablation campaign of 20+ model versions under identical 5-fold stratified cross-validation, holding the pipeline fixed while swapping backbones, losses, augmentations, training strategies, data scale, and label granularity, then reading off weighted-F1 and per-group F1.

Load-bearing premise

The big scaling gain and the 13-versus-45 label collapse are assumed to reflect real clinical tongue signal, even though much of the large merged set uses commercial-API, platform, or machine-generated labels with known noise rather than full expert annotation.

What would settle it

Retrain the same ConvNeXt-Tiny recipe on an independently collected multi-centre set that is fully expert-annotated for both the 13-label and 45-label schemes; if the ~0.78 weighted-F1 and the collapse to ~0.22 disappear or reverse under clean expert labels, the scaling and label-threshold claims fail.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Practitioners should prefer compact binary label sets with enough positives over fine-grained taxonomies that leave many classes nearly empty.
  • For color-dependent medical images, aggressive hue/saturation jitter should be treated as harmful by default, not as free regularization.
  • Weak tasks in a multi-head model are better fixed by dedicated single-group specialists than by averaging correlated full models.
  • On roughly thousand-image regimes, extra architectural complexity (multi-branch regions, contrastive pretraining, heavy attention pooling) is more likely to hurt than help.
  • Annotation budget is better spent on more samples with simpler reliable labels than on fewer samples with ambitious syndrome-level labels.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same label-complexity cliff should appear in other multi-label clinical tasks (skin findings, chest X-ray impressions, fundus signs) whenever positive counts per fine label fall below a few dozen.
  • If commercial or VLM-distilled tongue labels are systematically easier than expert judgment, published scaling curves may overstate deployable clinical accuracy until expert-only external tests are standard.
  • Specialist replacement for weak heads is a cheap ensemble pattern that could be tried wherever multi-task medical models show a few chronically failing labels amid strong ones.
  • A natural next measurement is the minimum positive-count per label at which expanding granularity stops being catastrophic under fixed total images.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The manuscript reports a large empirical ablation of deep multi-label classifiers for TCM tongue diagnosis across 20+ model versions under shared 5-fold CV. On the 976-image expert subset it compares backbones (ResNet50, ConvNeXt-Tiny/Small, EfficientNetV2-S, Swin-Tiny, DINOv2+LoRA), losses (BCE+pos_weight, Focal, ASL, LDAM-DRW), augmentations, and training tricks, peaking at weighted-F1 0.6625 with ConvNeXt-Tiny, restrained HSV jitter, and weak-group specialist replacement. On a merged 11,101-image set with 13 binary labels it reports weighted-F1 0.7761, and a 45-dimensional label expansion collapses to ~0.22. From this the authors distill six design principles claimed to generalize to imbalanced multi-label medical imaging.

Significance. If the controlled 976-sample findings hold, the paper supplies unusually thorough, reproducible guidance for a niche but clinically relevant multi-label setting: BCE+pos_weight over ASL on extreme imbalance, color-augmentation restraint when hue is diagnostic, parameter-efficient ConvNeXt-Tiny over larger CNNs/ViTs at small N, and specialist replacement beating probability averaging. Documented negative results (RegionMultiBranch, SupCon, curriculum, aggressive MixUp+pseudo-labels) are genuinely useful. The scaling and 13-vs-45 principles would be high-impact if cleanly isolated, but as written they are the least controlled claims. Strengths include fixed seed, shared folds, per-group breakdowns, std across folds, and explicit failure analysis.

major comments (4)
  1. [§4.6, Abstract, §5.1 P5] §4.6 and Abstract principle (5): the reported +20.6% (0.6433→0.7761) is not an isolated data-scaling ablation. v18 simultaneously (i) multiplies N ~11×, (ii) switches from 11-group/38-category expert multi-class heads (softmax+LDAM-DRW) to 13 independent binary sigmoid heads, (iii) drops LDAM/DRW, and (iv) merges TonguExpert (~54% of data) with platform/VLM/CV labels, zero positives for 6/13 labels, and stated 10–15% noise (§3.1, Table 1, Limitations). The headline scaling principle and its claimed clinical generalizability therefore confound N, label schema, head/loss design, and label source. At minimum, restate P5 as a joint N+schema result; ideally add a controlled bridge (e.g., 13-binary heads on the 976 expert set and/or on full TongueDx2 API labels alone) so the N effect can be separated from task reformulation.
  2. [§4.7, Abstract, §5.1 P6] §4.7 and Abstract principle (6): the 13→45 collapse (0.7761→0.2194) mixes extreme per-label sparsity with the same noisy pseudo-label pipeline used for TonguExpert and for many fine-grained dimensions. The manuscript already notes <50 positives and high distillation error for rare syndrome labels, so the failure cannot be attributed primarily to “label dimension count” as a clean design threshold. Please either (a) report a 45-d run restricted to expert-quality labels / high-prevalence subset, or (b) reframe P6 as “fine-grained + sparse + noisy labels fail,” and avoid presenting 13 vs 45 as a pure granularity threshold transferable to expert annotation settings.
  3. [Abstract, §5.1, Conclusions] Generalization claim (Abstract, Conclusions, §5.1): principles (1)–(4) are better controlled on the 976 expert ladder and support cautious transfer advice; principles (5)–(6) rest on non-expert majority labels and confounded design changes. The assertion that all six principles “are generalizable to multi-label medical image classification with class imbalance” overreaches the evidence, especially given single-institution data and no external clinical validation (Limitations). Narrow the claim: state which principles are supported by matched expert-label ablations versus which are observational on the merged API/VLM corpus, and temper clinical-transfer language accordingly.
  4. [§4.2, Table 4] §4.2 ASL vs BCE: the +2.7 wF1 result is load-bearing for P2 and is cleanly matched (v3 vs v3.5), which is a strength. The defense that γ_neg search is unnecessary because the failure is “mechanistic” is plausible but still incomplete for a methods paper: a short γ_neg∈{1,2,3,4} sweep (or ASL with the paper’s recommended medical-imbalance settings) on the same folds would close the main alternative explanation that the chosen γ_neg=4 was simply mistuned. This is a moderate request given how strongly P2 is marketed.
minor comments (6)
  1. [Abstract, §4.1, Table 3] Table 3 lists DINOv2+LoRA (v17) at 0.6501 as best single model, yet the abstract/headline 976-sample best is 0.6625 (ensemble). Clarify single-model vs ensemble SOTA consistently in abstract, §4.1, and figure captions.
  2. [§4.4, Table 7] §3.7 and §4.4: several “advanced strategies” are informative negatives; consider a compact summary table of what was held fixed vs changed for v10/v11/v12/v16-B2/B3 to aid replication.
  3. [§3.2] Mask-crop_v1→v2 gain (+8.6 pp, §3.2) is striking; state whether all reported v1–v17 numbers use v2 only, so readers do not mix preprocessing eras.
  4. [Figure 3] Figure 3 is described but numerical version IDs on the x-axis and explicit marking of confounded (multi-change) versions would improve readability.
  5. [§6 Conclusion] Minor wording: “TongueNet” appears in the Conclusion as if naming this system, while Related Work treats TongueNet as prior work [6]; disambiguate naming.
  6. [§3.5] Report exact pos_weight formula application per label and whether weights were recomputed per fold (methods reproducibility).

Circularity Check

0 steps flagged

No significant circularity: empirical 5-fold ablation bake-off; metrics and held-out comparisons are not forced by construction from inputs.

full rationale

This paper is a controlled empirical ablation study (20+ model versions, matched 5-fold stratified CV, weighted-F1 and per-group F1), not a first-principles derivation. Backbone, loss, augmentation, and training-strategy comparisons hold other factors fixed and report held-out fold metrics that are external to the training objective (BCE/LDAM/etc.). Data-scaling and 13-vs-45 label results are likewise measured outcomes on defined datasets, not quantities algebraically identical to fitted constants. Use of ImageNet/DINOv2 pretraining and commercial-API/VLM soft or pseudo labels is ordinary transfer/semi-supervised practice and does not redefine the evaluation target. Self-citations (e.g., TongueNet-related prior work) are background, not load-bearing uniqueness theorems that force the six design principles. Confounding of N with label schema and non-expert label noise affects causal isolation and generalizability, but that is a validity concern, not circular reduction of claim to input. No self-definitional loop, fitted-input-as-prediction, or ansatz-via-self-citation chain is present. Score 0; steps empty.

Axiom & Free-Parameter Ledger

8 free parameters · 6 axioms · 2 invented entities

Load-bearing content is experimental protocol plus domain labeling assumptions, not new physical entities. Claims rest on standard DL transfer assumptions, TCM label ontologies, and many hand-chosen hyperparameters that define each ‘version.’ Generalization beyond this image distribution is asserted rather than measured externally.

free parameters (8)
  • distillation weight λ_distill = 0.15
    Set to 0.15 (and earlier 0.3) by hand; co-activation with MixUp is reported catastrophic, so the chosen weight is part of the successful recipe.
  • HSV jitter magnitudes (restrained) = hue±0.01, sat/bright±0.05
    Hue/sat/brightness caps (0.01/0.05/0.05) chosen after aggressive settings failed; central to principle P3.
  • MixUp/CutMix schedule = α≈0.2, p≈0.2, λ_min=0.7, 70% CutMix
    α, probability, min λ≥0.7, and 70:30 CutMix:MixUp bias selected to protect color labels.
  • BCE positive class weights w_c = w_c=(N−/N+_c) per class
    Frequency-derived pos_weights are data-dependent rebalancing essential to the BCE-beats-ASL claim.
  • LDAM-DRW schedule and margins = DRW@epoch40, scale s=30, maxm=0.5
    Margin scale and DRW start epoch (40) are chosen hyperparameters for softmax groups on the 976-set.
  • Input resolution 320 and RandAugment M = 320px; M=5 (small-N), M=9 (11k)
    320 chosen vs 224/384 by validation; M=5 vs 9 ablated—part of reported optima.
  • LoRA rank/alpha and DINOv2 setup = r=16, α=32
    r=16, α=32, 336 input for v17; affects ‘best single model’ comparison.
  • Pseudo-label confidence threshold = 0.85
    0.85 threshold in failed v4; defines the negative pseudo-label result.
axioms (6)
  • domain assumption ImageNet or DINOv2 pretrained features transfer usefully to clinical tongue photographs after fine-tuning/LoRA.
    All backbone comparisons start from these pretrained weights (§3.4); without transfer the small-N results would not hold.
  • domain assumption TCM tongue attributes can be treated as the stated group-softmax and/or independent binary label systems with clinical meaning preserved under coarsening to 13 labels.
    Systems A/B/C in §3.3 define the prediction targets; the scaling and collapse claims are claims about these ontologies.
  • ad hoc to paper Commercial API, TonguExpert platform, and VLM/CV distilled labels are accurate enough to rank training methods and measure scaling.
    Majority of 11k supervision is non-expert (§3.1, §4.7); authors estimate 10–15% noise yet still treat wF1 gains as method truth.
  • domain assumption 5-fold stratified CV within the same collection environment unbiasedly estimates generalization for the stated design principles.
    No external test set; Limitations admit single-institution acquisition (§5.3).
  • domain assumption Weighted F1 (and per-group F1) is the right primary utility measure for comparing designs under imbalance.
    All principles are ranked by wF1 (§3.8); different metrics could reorder rare-class methods.
  • standard math Standard optimization mathematics (AdamW, AMP, EMA, cosine schedules) behaves as usual; reported deltas are due to the ablated factor.
    Implicit in the controlled-version protocol (§3–4).
invented entities (2)
  • TongueDx2 expert/API label resource and merged_v1 13-label corpus no independent evidence
    purpose: Provide the primary experimental substrate for ablations and scaling claims.
    Dataset construction and dual annotation paths are paper-specific resources, not a new physical entity; independent public release is incomplete (upon request).
  • Weak-group independent classifier replacement ensemble (B1) no independent evidence
    purpose: Boost G1/G2/G7/G8 by swapping in single-group specialists at inference.
    Methodological construct demonstrated only on this run suite; related to known specialist/ensemble ideas but specified here as targeted replacement vs averaging.

pith-pipeline@v1.2.0-daily-grok45 · 24899 in / 4574 out tokens · 95103 ms · 2026-07-31T16:23:00.684979+00:00 · methodology

0 comments
read the original abstract

Deep learning has shown promise for automated tongue diagnosis in traditional Chinese medicine (TCM), yet the design space remains underexplored. We conducted a systematic ablation study spanning 20+ model versions under rigorous 5-fold cross-validation on TongueDx2 (5,109 images, 976 expert-annotated) and a merged dataset of 11,101 samples. We compared six backbone architectures, four loss functions, five augmentation strategies, and six training strategies. The best 976-sample model achieved weighted-F1 of 0.6625 using ConvNeXt-Tiny with restrained augmentation and weak-group ensemble, while the best 11,101-sample model reached weighted-F1 of 0.7761. Six key design principles emerged: (1) ConvNeXt-Tiny offers optimal parameter efficiency; (2) BCE substantially outperforms Asymmetric Loss (+2.7%); (3) restrained color augmentation is critical; (4) weak-group ensemble replacement (+2.1%) outperforms probability averaging; (5) data scaling yielded +20.6% improvement; (6) expanding from 13 to 45 label dimensions caused catastrophic collapse (0.78 to 0.22). These principles are generalizable to multi-label medical image classification with class imbalance.

Figures

Figures reproduced from arXiv: 2607.28148 by Hanqing Zhao, Junze Geng, Linan Wang, Longxia Gao, Meng Zhang, Yuhe Han.

Figure 1
Figure 1. Figure 1: Overall pipeline architecture. The system follows a multi-stage design: YOLO-based [PITH_FULL_IMAGE:figures/full_fig_p018_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Dataset statistics. (A) Label distribution across 11 groups for TongueDx2 (976 expert [PITH_FULL_IMAGE:figures/full_fig_p019_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Experimental results overview. Weighted-F1 scores across all 20+ model versions on [PITH_FULL_IMAGE:figures/full_fig_p019_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Backbone architecture comparison. Mean weighted-F1 ( [PITH_FULL_IMAGE:figures/full_fig_p020_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Loss function ablation heatmap. Per-group F1 scores for four loss function variants (v3 [PITH_FULL_IMAGE:figures/full_fig_p021_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Impact of data scaling on per-label performance. (Left) v15a on 976 samples (wF1 [PITH_FULL_IMAGE:figures/full_fig_p021_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Label design impact. Weighted-F1 comparison: 13-dimension (v18 [PITH_FULL_IMAGE:figures/full_fig_p022_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Weak-group ensemble analysis. Per-group F1 comparison between v15a single model [PITH_FULL_IMAGE:figures/full_fig_p023_8.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

31 extracted references · 6 canonical work pages

  1. [1]

    He, K., Zhang, X., Ren, S., &Sun, J.(2016).Deepresiduallearningforimagerecognition.In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 770–778. doi:10.1109/CVPR.2016.90

    doi:10.1109/cvpr.2016.90 2016

  2. [2]

    Liu, Z., Mao, H., Wu, C.-Y., Feichtenhofer, C., Darrell, T., & Xie, S. (2022). A ConvNeXt forthe2020s. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 11976–11986. doi:10.1109/CVPR52688.2022.01167 23

    arXiv 2022

  3. [3]

    Tan, M., &Le, Q.(2021).EfficientNetV2: Smallermodelsandfastertraining.InProceedings of the 38th International Conference on Machine Learning (ICML)

    2021

  4. [4]

    Oquab, M., Darcet, T., Moutakanni, T., et al. (2024). DINOv2: Learning robust visual features without supervision.Transactions on Machine Learning Research (TMLR)

    2024

  5. [5]

    J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S.,

    Hu, E. J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., ... & Chen, W. (2022). LoRA: Low-rank adaptation of large language models. InInternational Conference on Learning Representations (ICLR)

    2022

  6. [6]

    Yang, L., Zhang, Y., Zhu, Y., et al. (2025). TongueNet: a multi-modal fusion and multi- label classification model for traditional Chinese Medicine tongue diagnosis.Frontiers in Physiology, 16, 1527751. doi:10.3389/fphys.2025.1527751

    arXiv 2025

  7. [7]

    Zhang, B., Li, X., & Zhang, D. (2013). Tongue color analysis for medical ap- plication.Evidence-Based Complementary and Alternative Medicine, 2013, 264742. doi:10.1155/2013/264742

    work page doi:10.1155/2013/264742 2013

  8. [8]

    Wang, X., et al. (2020). Using deep convolutional neural network for recognizing unhealthy tongue.Informatics in Medicine Unlocked, 20, 100388

    2020

  9. [9]

    Jiang, T., et al. (2021). Tongue image quality assessment based on a deep convolu- tional neural network.BMC Medical Informatics and Decision Making, 21(Suppl 3), 150. doi:10.1186/s12911-021-01508-8

    work page doi:10.1186/s12911-021-01508-8 2021

  10. [10]

    Liu, Q., et al. (2023). A survey of artificial intelligence in tongue image for disease diagnosis. Healthcare, 11(16), 2258

    2023

  11. [11]

    Irvin, J., Rajpurkar, P., Ko, M., Yu, Y., Ciurea-Ilcus, S., Chute, C., ... & Ng, A. (2019). CheXpert: A large chest radiograph dataset with uncertainty labels and expert compar- ison. InProceedings of the AAAI Conference on Artificial Intelligence, 33(01), 590–597. doi:10.1609/aaai.v33i01.3301590

    doi:10.1609/aaai.v33i01.3301590 2019

  12. [12]

    A., Ko, J., Swetter, S

    Esteva, A., Kuprel, B., Novoa, R. A., Ko, J., Swetter, S. M., Blau, H. M., & Thrun, S. (2017). Dermatologist-level classification of skin cancer with deep neural networks.Nature, 542(7639), 115–118. doi:10.1038/nature21056

    doi:10.1038/nature21056 2017

  13. [13]

    C., Wu, D., Narayanaswamy, A.,

    Gulshan, V., Peng, L., Coram, M., Stumpe, M. C., Wu, D., Narayanaswamy, A., ... & Webster, D. R. (2016). Development and validation of a deep learning algorithm for de- tection of diabetic retinopathy in retinal fundus photographs.JAMA, 316(22), 2402–2410. doi:10.1001/jama.2016.17216

    arXiv 2016

  14. [14]

    M., & de la Iglesia-Vaya, M

    Bustos, A., Pertusa, A., Salinas, J. M., & de la Iglesia-Vaya, M. (2020). PadChest: A large chest X-ray image dataset with multi-label annotated reports.Applied Sciences, 10(21),

    2020

  15. [15]

    Cao, K., Wei, C., Gaidon, A., Arechiga, N., & Ma, T. (2019). Learning imbalanced datasets with label-distribution-aware margin loss. InAdvances in Neural Information Processing Systems (NeurIPS), 32

    2019

  16. [16]

    Lin, T.-Y., Goyal, P., Girshick, R., He, K., & Dollár, P. (2017). Focal loss for dense ob- ject detection. InProceedings of the IEEE International Conference on Computer Vision (ICCV), pp. 2980–2988. doi:10.1109/ICCV.2017.324

    doi:10.1109/iccv.2017.324 2017

  17. [17]

    Ridnik, T., Ben-Baruch, E., Zamir, N., Noy, A., Friedman, I., Protter, M., & Zelnik-Manor, L. (2021). Asymmetric loss for multi-label classification. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pp. 82–91. doi:10.1109/ICCV48522.2021.00087 24

    arXiv 2021

  18. [18]

    Hinton, G., Vinyals, O., & Dean, J. (2015). Distilling the knowledge in a neural network. arXiv preprintarXiv:1503.02531

    Pith/arXiv arXiv 2015

  19. [19]

    Lee, D.-H. (2013). Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. InWorkshop on Challenges in Representation Learning, ICML

    2013

  20. [20]

    &Krishnan, D.(2020)

    Khosla, P., Teterwak, P., Wang, C., Sarna, A., Tian, Y., Isola, P., ... &Krishnan, D.(2020). Supervised contrastive learning. InAdvances in Neural Information Processing Systems (NeurIPS), 33

    2020

  21. [21]

    N., & Lopez-Paz, D

    Zhang, H., Cissé, M., Dauphin, Y. N., & Lopez-Paz, D. (2018). mixup: Beyond empirical risk minimization. InInternational Conference on Learning Representations (ICLR)

    2018

  22. [22]

    J., Chun, S., Choe, J., & Yoo, Y

    Yun, S., Han, D., Oh, S. J., Chun, S., Choe, J., & Yoo, Y. (2019). CutMix: Regu- larization strategy to train strong classifiers with localizable features. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pp. 6023–6032. doi:10.1109/ICCV.2019.00612

    arXiv 2019

  23. [23]

    D., Zoph, B., Shlens, J., & Le, Q

    Cubuk, E. D., Zoph, B., Shlens, J., & Le, Q. V. (2020). RandAugment: Practical auto- mated data augmentation with a reduced search space. InAdvances in Neural Information Processing Systems (NeurIPS), 33

    2020

  24. [24]

    & Guo, B

    Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., ... & Guo, B. (2021). Swin transformer: Hierarchical vision transformer using shifted windows. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pp. 10012–10022. doi:10.1109/ICCV48922.2021.00986

    arXiv 2021

  25. [25]

    Li, T., et al. (2025). TonguExpert: A deep learning-based algorithm platform for fine- grained extraction and classification of tongue phenotypes.Phenomics, 5

    2025

  26. [26]

    Jiang, T., Lu, Z., Hu, X., et al. (2022). Deep learning multi-label tongue image analysis and its application in a population undergoing routine medical checkup.Evidence-Based Complementary and Alternative Medicine, 2022, 3384209. doi:10.1155/2022/3384209

    work page doi:10.1155/2022/3384209 2022

  27. [27]

    Liu, H., et al. (2024). Research on multi-label recognition of tongue features in stroke pa- tients based on deep learning.Scientific Reports, 14, 84002. doi:10.1038/s41598-024-84002-1

    work page doi:10.1038/s41598-024-84002-1 2024

  28. [28]

    Peng, C., et al. (2026). Application of a large visual language model on tongue image de- scription generation and physical constitution reasoning (TongueVLM): Model development and validation study.JMIR Medical Informatics, 4, e87237. doi:10.2196/87237

    work page doi:10.2196/87237 2026

  29. [29]

    doi:10.1136/bmj-2023-078378

    Collins, G.S., etal.(2024).TRIPOD+AIstatement: updatedguidanceforreportingclinical prediction models that use regression or machine learning methods.BMJ, 385, e078378. doi:10.1136/bmj-2023-078378

    doi:10.1136/bmj-2023-078378 2024

  30. [30]

    S., et al

    Tejani, A. S., et al. (2024). CLAIM 2024: A checklist for artificial intelligence in medical imaging.Radiology: Artificial Intelligence, 6(3), 240300. doi:10.1148/ryai.240300 25

    doi:10.1148/ryai.240300 2024

  31. [7135]

    doi:10.3390/app10217135

    work page doi:10.3390/app10217135