Pith. sign in

REVIEW 3 major objections 5 minor 29 references

Calibrated Selective Prediction Using Deep Ensembles for ROI-Based Thyroid Nodule Ultrasound Classification Under Dataset Shift: A Retrospective Evaluation

T0 review · 3 major / 5 minor · reviewed 2026-07-15 · grok-4.5

Pith's one-line read A calibrated deep ensemble can support conservative selective triage of thyroid ultrasound images, but its frozen operating thresholds do not transport under dataset shift.

desk verdict Careful empirical stress test of a frozen calibrated ensemble triage policy: strong internal numbers, honest external failure, no clinical overclaim. read the letter →

arxiv 2607.12075 v1 pith:WZWWWITL submitted 2026-07-13 eess.IV cs.AIcs.CV

classification eess.IVcs.AIcs.CV
keywords thyroidnoduleultrasounddeeplearningselectivepredictionuncertaintyestimationcalibrationdatasetshiftfine-needleaspiration
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 asks whether a carefully calibrated five-model deep ensemble, using disagreement among the models as a deferral signal, can safely issue image-based triage suggestions for thyroid nodules on ultrasound. Internally, on a large malignant-enriched public set, the system is highly discriminative and well calibrated, and a strict three-tier policy (no-biopsy suggestion, biopsy recommendation, or radiologist review) keeps almost all cancers out of the automated no-biopsy path. Externally, on a second public set with different prevalence, annotations, and ROI construction, discrimination falls, calibration collapses, and the frozen rules send most images to review while the biopsy-path predictive value drops. A sympathetic reader cares because high accuracy alone is not enough for clinical decision support; the work shows both that selective prediction can withhold unreliable suggestions and that operating thresholds themselves are fragile under real dataset shift. The authors therefore present selective prediction as a cautious routing mechanism, not as evidence of deployable biopsy reduction.

What carries the argument

The three-tier MI-guided triage policy: mutual information among five calibrated ensemble members first defers high-disagreement images; remaining low-disagreement cases receive a No-FNA suggestion only if mean probability is below a high-NPV rule-out threshold, an FNA recommendation only if above a high-PPV rule-in threshold, and otherwise also go to review. This turns ensemble disagreement into an operational deferral score that withholds automated suggestions when they are least reliable.

What would settle it

If local vector-scaling recalibration and re-derivation of the MI and probability thresholds on a modest site-specific validation set drawn from the same distribution as the external cohort recovered at least 95 percent FNA PPV and 98 percent No-FNA NPV with autonomous coverage comparable to the internal rates, the limited-transportability claim would be overturned; if those targets remained unreachable without near-total deferral, the claim would be confirmed.

Watch

Extended reading notes

Core claim

On the development dataset the calibrated five-member ConvNeXt-Tiny ensemble reaches pooled out-of-fold AUC-ROC 0.9395 and expected calibration error 0.0088; a mutual-information gate plus probability thresholds can issue autonomous no-FNA and FNA suggestions while capturing 99.83 percent of malignancies. When the identical frozen model and thresholds are applied to an independent external cohort, discrimination and calibration both degrade and the policy largely collapses into radiologist review, showing that selective prediction can flag unsuitable images but that the operating policy itself does not transport without local recalibration.

Load-bearing premise

The paper assumes that image-level splits of the development set, made without patient identifiers, and lesion crops taken from provided boxes or masks are independent and representative enough that the internal thresholds remain a fair test of external transportability.

Editorial extensions

If this is right

  • Autonomous image-based No-FNA and FNA suggestions should be issued only for low-disagreement, probability-extreme cases; intermediate or high-disagreement images should be routed to human review.
  • Strong internal discrimination and calibration do not guarantee that probability or uncertainty thresholds remain valid under dataset shift.
  • Local recalibration of probabilities and re-validation of MI and decision thresholds are required before any frozen policy can be considered at a new site.
  • Selective prediction can preserve malignancy capture mainly by increasing the review burden when the data distribution changes.
  • Claims of actual biopsy reduction cannot be made from ROI-level retrospective assignment alone; clinical variables and prospective workflow studies remain necessary.

Reading between the lines

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

  • The jump to 83.7 percent external review implies that the uncertainty score itself is distribution-sensitive, so fixed MI cut-offs may need to be replaced by domain-adaptive or conformal deferral rules.
  • Because internal splits were image-level without patient IDs, residual same-patient correlation could have inflated the internal calibration and selective-risk numbers against which transportability is judged.
  • MSP-derived uncertainty ranked internal errors better than MI; testing which score retains better external ranking would clarify whether epistemic disagreement or pure confidence is the more portable safety signal.
  • The same selective-prediction template could be stress-tested on other ultrasound tasks where annotation style and prevalence also shift between public cohorts.
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

3 major / 5 minor

Summary. The manuscript develops a calibrated five-member deterministic ConvNeXt-Tiny ensemble (with SE attention) for annotation-assisted ROI-based thyroid-nodule ultrasound classification and selective image-based triage. TN5000 is used for five-fold CV, member-wise vector-scaling calibration on held-out val_calib, and fold-specific MI/probability threshold selection on val_es; a frozen deployment model and 50% nominal-MI policy are then applied unchanged to the independent TN3K cohort. Internally the ensemble reports strong pooled OOF discrimination and calibration (AUC-ROC 0.9395, AP 0.9715, ECE 0.0088) and a conservative three-tier policy (No-FNA / FNA / radiologist review) that retains high malignancy capture. Externally discrimination falls, calibration collapses, and the frozen thresholds produce an 83.7% review rate with degraded FNA PPV, supporting the authors’ claim of limited threshold transportability under dataset shift.

Significance. If the reported internal-to-external gap holds, the work supplies a carefully documented empirical demonstration that calibrated ensemble disagreement can support conservative selective referral on a malignant-enriched development set while the same frozen operating policy fails to transport. The clean separation of calibration, threshold derivation, and external evaluation, together with risk-coverage comparisons and explicit non-claims of clinical deployability or biopsy reduction, is a useful methodological contribution for medical imaging under dataset shift. Public datasets, detailed protocol, and code availability further strengthen reproducibility. The result is incremental rather than transformative, but it is a solid, cautious addition to the selective-prediction literature in thyroid ultrasound.

major comments (3)
  1. Sections 3.1 and 6 correctly flag that TN5000 patient identifiers are unavailable, so image-level stratified splits cannot guarantee patient-level independence. Residual patient-level leakage would inflate the internal OOF AUC/ECE and the internal-to-external gap that underpins the transportability claim. The authors already treat this as a limitation; a quantitative sensitivity analysis (e.g., approximate patient clustering if any metadata exist, or a leave-one-site-style ablation) or a clearer statement that the internal numbers are upper bounds would make the central comparison more robust.
  2. Section 3.6 and Supplementary S2 show that MSP-derived uncertainty and predictive variance both outperform MI on AURC/AUGRC (0.03020/0.02261 and 0.03234/0.02424 vs MI 0.03778/0.02722). The decision to retain MI for the frozen external policy is hypothesis-driven and transparent, yet the external transportability conclusion is therefore tied to a sub-optimal ranking score. Reporting the same frozen-policy experiment with the best internal score (or at least MSP) would test whether the observed TN3K failure is score-specific or general.
  3. Supplementary S1 documents that TN5000 ROIs are directly resized while TN3K ROIs use aspect-preserving resize plus black padding; annotation type also differs (boxes vs masks). These preprocessing differences are part of the declared dataset shift, but their contribution to the ECE rise (0.0088 → 0.1899) and MI-gate failure (only 20.5% of TN3K pass the frozen τ_MI) is not isolated. A controlled re-processing of one dataset with the other’s pipeline would strengthen the claim that the degradation is domain shift rather than ROI-construction artifact.
minor comments (5)
  1. Abstract and Results report slightly different pooled AP values (0.9715 vs fold-mean 0.9720); a single consistent figure should be used throughout.
  2. Table 7 and the surrounding text use “nominal MI retention”; a one-sentence reminder that this is the val_es MI-gate fraction, not final autonomous coverage, would reduce reader confusion.
  3. Figure 5 Grad-CAM examples are qualitative only; the caption already states this, but the main text could more explicitly note that they are single-member (seed 42) maps and not used for any decision.
  4. A few typographical inconsistencies appear (e.g., “au tomated”, “multicentre decision-support”, missing spaces after commas in places); a final copy-edit pass is warranted.
  5. The web prototype (Supplementary Figure S4) is a nice demonstration but is not evaluated; a single sentence clarifying that it is outside the scientific claims would prevent over-interpretation.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: empirical hold-out evaluation of a fitted selective-triage policy with honest external degradation.

full rationale

The paper is a standard retrospective ML evaluation. Models, member-wise vector-scaling parameters, MI thresholds, and probability rule-out/rule-in thresholds are fitted exclusively on TN5000 train/val_es/val_calib partitions (Sections 3.4–3.10); held-out fold and pooled OOF metrics, plus the frozen-policy application to TN3K, never re-enter fitting. External TN3K labels are used only for outcome reporting. The three-tier policy and NPV/PPV targets are operational choices evaluated after the fact, not definitions that force the reported discrimination, calibration, or transportability numbers. No self-definitional equations, no fitted quantity renamed as an independent prediction, no load-bearing self-citation uniqueness claims, and no ansatz smuggled via prior author work appear. The central claim (strong internal calibrated selective performance + limited frozen-threshold transportability) is therefore self-contained against the reported benchmarks.

Assumptions & free parameters 4 free parameters · 5 assumptions · 1 invented entities

The work rests on standard deep-learning and selective-classification machinery plus a small set of operational targets chosen by the authors. No new physical entities are postulated; free parameters are the usual fitted calibration and threshold values plus architecture choices. Domain assumptions about ROI quality and image-level independence are the main load-bearing premises that are not independently verified.

free parameters (4)
  • vector-scaling weights and biases (per member, per fold)
    Fitted by cross-entropy minimization on val_calib only; directly determine calibrated probabilities used for triage.
  • MI retention grid and resulting τ_MI, τ_low, τ_high
    Chosen on val_es to meet 98% NPV / 95% PPV operational targets; frozen for external evaluation.
  • ensemble size M=5 and random seeds
    Hand-chosen; controls the disagreement score that drives deferral.
  • backbone and SE module choice
    Selected by preliminary validation AUC screen on the same source data.
assumptions (5)
  • domain assumption Image-level stratified splits of TN5000 yield sufficiently independent train/val/test estimates despite missing patient IDs.
    Stated in Sections 3.1 and 6; residual patient correlation would inflate internal metrics.
  • domain assumption Provided bounding boxes / masks produce ROIs representative of clinical lesion appearance.
    Annotation-assisted design (S1); localization error is not modeled.
  • domain assumption Mutual information of calibrated member probabilities is a usable operational proxy for epistemic disagreement.
    Adopted as primary deferral score even though MSP ranked better internally (S2).
  • ad hoc to paper 98% NPV and 95% PPV are appropriate retrospective safety targets for image-based triage.
    Explicitly described as operational choices, not clinical standards (3.9).
  • standard math Standard selective-classification risk-coverage evaluation and deep-ensemble uncertainty theory apply.
    Citations to Geifman, Lakshminarayanan, Ovadia, etc.
invented entities (1)
  • strict three-tier MI-guided image-based triage policy (No-FNA / FNA / review)
    purpose: Map calibrated probability and ensemble disagreement onto retrospective clinical-style pathways without claiming actual biopsy reduction.
    Policy construct introduced by the authors; evaluated only retrospectively on two public datasets.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Calibrated Selective Prediction Using Deep Ensembles for ROI-Based Thyroid Nodule Ultrasound Classification Under Dataset Shift: A Retrospective Evaluation." pith.science (2026). https://pith.science/paper/WZWWWITL

@misc{pith2026260712075,
  author       = {Pith},
  title        = {Pith review of: Calibrated Selective Prediction Using Deep Ensembles for ROI-Based Thyroid Nodule Ultrasound Classification Under Dataset Shift: A Retrospective Evaluation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WZWWWITL}},
  note         = {Machine review of arXiv:2607.12075}
}
read the original abstract

Background: Deep learning models can classify thyroid nodules on ultrasound, but reliable clinical decision support also requires calibrated probabilities, uncertainty estimation, and selective referral, particularly under dataset shift. Methods: We developed a calibrated deterministic five-member deep ensemble for ROI-based thyroid nodule classification and selective image-based triage. TN5000 was used for model development, five-fold cross-validation, member-wise vector-scaling calibration, and fold-specific threshold selection. TN3K served as an independent external dataset-shift evaluation. The framework used ConvNeXt-Tiny with squeeze-and-excitation attention, ensemble-mean malignancy probability, and mutual information (MI) as an ensemble-disagreement score. A three-tier policy assigned images to No-FNA suggestion, FNA recommendation, or radiologist review. Results: On pooled out-of-fold TN5000 predictions, the ensemble achieved AUC-ROC 0.9395, AP 0.9715, ECE 0.0088, and Brier score 0.0813. At 50% nominal MI retention, 7.2% of cases received a No-FNA suggestion, 39.9% an FNA recommendation, and 52.9% radiologist review, with 98.3% No-FNA NPV and 99.83% malignancy capture. On TN3K, AUC-ROC decreased to 0.7870, AP to 0.7254, ECE increased to 0.1899, and Brier score to 0.2281. The frozen TN5000 policy assigned 83.7% to review, 1.0% to No-FNA, and 15.3% to FNA recommendation. No malignant image entered the No-FNA pathway, but FNA-recommendation PPV fell to 76.6%. Conclusion: The framework showed strong internal discrimination and calibration, but limited external threshold transportability. Selective prediction may help identify images unsuitable for automated triage, but local recalibration, threshold validation, and prospective clinical evaluation are required before deployment.

Figures

Figures reproduced from arXiv: 2607.12075 by the authors.

Figure 1
Figure 1. Overview of the annotation-assisted selective-prediction workflow. 3.3 Backbone Selection and Final Model Architecture A preliminary development screen compared ConvNeXt-Tiny, EfficientNetV2-S, DenseNet121, ResNet18, and MobileNetV4 using the same development protocol. Validation AUC-ROC was the primary criterion for backbone selection because it summarizes discrimination across classification thresholds. Average pr… view at source ↗
Figure 2
Figure 2. Pooled out-of-fold ROC and precision-recall curves for the calibrated five-member ConvNeXt-Tiny ensemble on TN5000 [PITH_FULL_IMAGE:figures/full_fig_p018_2.png] view at source ↗
Figure 3
Figure 3. Pooled out-of-fold reliability diagram of calibrated ensemble predictions on TN5000. 4.3.2 Selective-Prediction Performance Selective-prediction performance was evaluated using pooled OOF predictions ( [PITH_FULL_IMAGE:figures/full_fig_p018_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Pooled out-of-fold MI risk-coverage curve. Cases were retained in ascending order of mutual information. Lower selective risk indicates fewer classification errors among retained images. The curve was calculated at all 5,000 retained-set sizes. 4.4 MI-Guided Strict Tri…
Figure 5
Figure 5. Figure 5: Qualitative Grad-CAM examples showing a true benign, true malignant, misclassified, and high-MI ROI. Rows show the original ROI and Grad-CAM overlay. 4.6 External Discrimination and Calibration Under Dataset Shift The frozen five-member deployment ensemble was applied …
Figure 6
Figure 6. Figure 6: AUC-ROC and Precision-Recall (PR) curves in external TN3K evaluation [PITH_FULL_IMAGE:figures/full_fig_p022_6.png]
Figure 7
Figure 7. Figure 7: External discrimination and fixed-threshold classification performance on TN3K. 4.7 External MI-Guided Triage Under Dataset Shift The fixed 50% nominal MI-retention policy, including member-specific vector scaling and triage thresholds derived exclusively from the TN50…
Figure 8
Figure 8. Figure 8: External strict MI-guided triage assignments on TN3K using fixed TN5000-derived thresholds. 5. Discussion This study evaluated whether a calibrated deterministic deep ensemble could support selective prediction for ROI-based thyroid nodule ultrasound classification by …

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 13 canonical work pages

  1. [1]

    B. R. Haugen et al., “2015 American Thyroid Association management guidelines for adult patients with thyroid nodules and differentiated thyroid cancer: The American Thyroid Association Guidelines Task Force on Thyroid Nodules and Differentiated Thyroid Cancer,” Thyroid, vol. 26, no. 1, pp. 1–133, 2016, doi: 10.1089/thy.2015.0020

  2. [2]

    ACR Thyroid Imaging, Reporting and Data System (TI-RADS): White paper of the ACR TI-RADS Committee,

    F. N. Tessler et al., “ACR Thyroid Imaging, Reporting and Data System (TI-RADS): White paper of the ACR TI-RADS Committee,” Journal of the American College of Radiology , vol. 14, no. 5, pp. 587–595, 2017, doi: 10.1016/j.jacr.2017.01.046

  3. [3]

    Deep learning based classification of ultrasound images for thyroid nodules: A large scale of pilot study,

    Q. Guan et al., “Deep learning based classification of ultrasound images for thyroid nodules: A large scale of pilot study,” Annals of Translational Medicine , vol. 7, no. 7, Art. no. 137, 2019, doi: 10.21037/atm.2019.04.34

  4. [4]

    Deep learning-based artificial intelligence model to assist thyroid nodule diagnosis and management: A multicentre diagnostic study,

    S. Peng et al., “Deep learning-based artificial intelligence model to assist thyroid nodule diagnosis and management: A multicentre diagnostic study,” The Lancet Digital Health, vol. 3, no. 4, pp. e250– e259, 2021, doi: 10.1016/S2589-7500(21)00041-8

  5. [5]

    Thyroid nodule classification in ultrasound imaging using deep transfer learning,

    Y. Xu, M. Xu, Z. Geng, J. Liu, and B. Meng, “Thyroid nodule classification in ultrasound imaging using deep transfer learning,” BMC Cancer, vol. 25, Art. no. 544, 2025, doi: 10.1186/s12885 -025- 13917-3

  6. [6]

    Deep learning technology for classification of thyroid nodules using multi -view ultrasound images: Potential benefits and challenges in clinical application,

    J. Kim et al ., “Deep learning technology for classification of thyroid nodules using multi -view ultrasound images: Potential benefits and challenges in clinical application,” Endocrinology and Metabolism, vol. 40, no. 2, pp. 216–224, 2025, doi: 10.3803/EnM.2024.2058

  7. [7]

    Deep learning for the diagnosis of suspicious thyroid nodules based on multimodal ultrasound images,

    Y. Tao et al., “Deep learning for the diagnosis of suspicious thyroid nodules based on multimodal ultrasound images,” Frontiers in Oncology , vol. 12, Art. no. 1012724, 2022, doi: 10.3389/fonc.2022.1012724

  8. [8]

    A multi-view deep learning model for thyroid nodules detection and characterization in ultrasound imaging,

    S. Vahdati, B. Khosravi, K. A. Robinson, P. Rouzrokh, M. Moassefi, Z. Akkus, and B. J. Erickson, “A multi-view deep learning model for thyroid nodules detection and characterization in ultrasound imaging,” Bioengineering, vol. 11, no. 7, Art. no. 648, 2024, doi: 10.3390/bioengineering11070648

Show all 29 references
  1. [9]

    From bench-to-bedside: How artificial intelligence is changing thyroid nodule diagnostics, a systematic review,

    V. R. Sant et al., “From bench-to-bedside: How artificial intelligence is changing thyroid nodule diagnostics, a systematic review,” Journal of Clinical Endocrinology and Metabolism, vol. 109, no. 7, pp. 1684–1693, 2024, doi: 10.1210/clinem/dgae277

  2. [10]

    Deep learning for classification of thyroid nodules on ultrasound: Validation on an independent dataset,

    J. Weng et al., “Deep learning for classification of thyroid nodules on ultrasound: Validation on an independent dataset,” Clinical Imaging , vol. 99, pp. 60 –66, 2023, doi : 10.1016/j.clinimag.2023.04.010

  3. [11]

    Optimizing thyroid nodule management with artificial intelligence: Multicenter retrospective study on reducing unnecessary fine needle aspirations,

    J.-H. Ni et al., “Optimizing thyroid nodule management with artificial intelligence: Multicenter retrospective study on reducing unnecessary fine needle aspirations,” JMIR Medical Informatics, vol. 13, Art. no. e71740, 2025, doi: 10.2196/71740

  4. [12]

    Enhancing automatic diagnosis of thyroid nodules from ultrasound scans leveraging deep learning models,

    A. Rashed, T. Medhat, and A. Elgarayhi, “Enhancing automatic diagnosis of thyroid nodules from ultrasound scans leveraging deep learning models,” Scientific Reports, vol. 15, Art. no. 40364, 2025, doi: 10.1038/s41598-025-25780-0

  5. [13]

    On calibration of modern neural networks,

    C. Guo, G. Pleiss, Y. Sun, and K. Q. Weinberger, “On calibration of modern neural networks,” in Proceedings of the 34th International Conference on Machine Learning, vol. 70, 2017, pp. 1321–1330

  6. [14]

    A review of uncertainty estimation and its application in medical imaging,

    K. Zou, Z. Chen, X. Yuan, X. Shen, M. Wang, and H. Fu, “A review of uncertainty estimation and its application in medical imaging,” Meta-Radiology, vol. 1, no. 1, Art. no. 100003, 2023, doi: 10.1016/j.metrad.2023.100003

  7. [15]

    Uncertainty-aware deep learning in healthcare: A scoping review,

    T. J. Loftus et al ., “Uncertainty-aware deep learning in healthcare: A scoping review,” PLOS Digital Health, vol. 1, no. 8, Art. no. e0000085, 2022, doi: 10.1371/journal.pdig.0000085

  8. [16]

    Simple and scalable predictive uncertainty estimation using deep ensembles,

    B. Lakshminarayanan, A. Pritzel, and C. Blundell, “Simple and scalable predictive uncertainty estimation using deep ensembles,” in Advances in Neural Information Processing Systems , vol. 30, 2017

  9. [17]

    Dropout as a Bayesian approximation: Representing model uncertainty in deep learning,

    Y. Gal and Z. Ghahramani, “Dropout as a Bayesian approximation: Representing model uncertainty in deep learning,” in Proceedings of the 33rd International Conference on Machine Learning, vol. 48, 2016, pp. 1050–1059

  10. [18]

    Can you trust your model’s uncertainty? Evaluating predictive uncertainty under dataset shift,

    Y. Ovadia et al., “Can you trust your model’s uncertainty? Evaluating predictive uncertainty under dataset shift,” in Advances in Neural Information Processing Systems, vol. 32, 2019

  11. [19]

    Role of calibration in uncertainty-based referral for deep learning,

    R. Zhang, C. Gatsonis, and J. A. Steingrimsson, “Role of calibration in uncertainty-based referral for deep learning,” Statistical Methods in Medical Research , vol. 32, no. 5, pp. 927 –943, 2023, doi : 10.1177/09622802231158811

  12. [20]

    Selective classification for deep neural networks,

    Y. Geifman and R. El -Yaniv, “Selective classification for deep neural networks,” Advances in Neural Information Processing Systems, vol. 30, 2017

  13. [21]

    Overcoming common flaws in the evaluation of selective classification systems,

    J. Traub, T. J. Bungert, C. T. Lüth, M. Baumgartner, K. H. Maier -Hein, L. Maier-Hein, and P. F. Jäger, “Overcoming common flaws in the evaluation of selective classification systems,” arXiv:2407.01032, 2024

  14. [22]

    TN5000: An ultrasound image dataset for thyroid nodule detection and classification,

    H. Zhang, Q. Liu, X. Han, L. Niu, and W. Sun, “TN5000: An ultrasound image dataset for thyroid nodule detection and classification,” Scientific Data, vol. 12, Art. no. 1437, 2025, doi: 10.1038/s41597- 025-05757-4

  15. [23]

    An ultrasonography of thyroid nodules dataset with pathological diagnosis annotation for deep learning,

    X. Hou et al ., “An ultrasonography of thyroid nodules dataset with pathological diagnosis annotation for deep learning,” Scientific Data, vol. 11, Art. no. 1272, 2024, doi: 10.1038/s41597 -024- 04156-5

  16. [24]

    Deep learning-based classification of thyroid nodules using uncertainty -aware multi-modal ultrasound imaging,

    M. Saini, T. A. Parvar, M. Velarde, N. B. Larson, M. Fatemi, and A. Alizad, “Deep learning-based classification of thyroid nodules using uncertainty -aware multi-modal ultrasound imaging,” Scientific Reports, vol. 16, Art. no. 4938, 2026, doi: 10.1038/s41598-026-35965-w

  17. [25]

    MTL -TN: An uncertainty -weighted multi-task framework for unified thyroid nodule segmentation and TI -RADS classification,

    T. Xiang, B. Li, Z. Hu, X. Wang, Y. Dang, and Y. Deng, “MTL -TN: An uncertainty -weighted multi-task framework for unified thyroid nodule segmentation and TI -RADS classification,” in Proc. 2025 5th Int. Conf. Computer, Internet of Things and Control Engineering (CITCE), 2025,...

  18. [26]

    Thyroid nodule and lymph node metastasis assessment from ultrasound images using deep learning,

    X. Zhao et al., “Thyroid nodule and lymph node metastasis assessment from ultrasound images using deep learning,” Frontiers in Neuroscience , vol. 19, Art. no. 1684104, 2025, doi: 10.3389/fnins.2025.1684104

  19. [27]

    A ConvNet for the 2020s,

    Z. Liu, H. Mao, C. -Y. Wu, C. Feichtenhofer, T. Darrell, and S. Xie, “A ConvNet for the 2020s,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2022, pp. 11966–11976, doi: 10.1109/CVPR52688.2022.01167

  20. [28]

    Squeeze-and-excitation networks,

    J. Hu, L. Shen, S. Albanie, G. Sun, and E. Wu, “Squeeze-and-excitation networks,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018, pp. 7132–7141, doi: 10.1109/CVPR.2018.00745

  21. [29]

    Grad-CAM: Visual explanations from deep networks via gradient -based localization,

    R. R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, and D. Batra, “Grad-CAM: Visual explanations from deep networks via gradient -based localization,” in Proceedings of the IEEE International Conference on Computer Vision, 2017, pp. 618–626, doi: 10.1109/ICCV.2017.74...

Pith tools

Reviewed July 15, 2026 · model on record in the stance chip above.