Pith. sign in

REVIEW 5 major objections 5 minor 29 references

Improving Generalization in MRI-Based Deep Learning Models for Total Knee Replacement Prediction

T0 review · 5 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read A knee-replacement MRI classifier trained on one scan protocol can generalize to another protocol without target labels by swapping batch normalization for instance normalization, adding intensity-preserving augmentation, and aligning…

desk verdict A useful empirical combination of IN + GIN + contrastive loss for TKR prediction, but the target model selection uses unlabeled target data, which weakens the single-source DG claim. read the letter →

arxiv 2504.19203 v7 pith:MW7OR37Z submitted 2025-04-27 eess.IV cs.CV

Ehsan Karami , Hamid Soltanian-Zadeh This is my paper · ORCID
classification eess.IVcs.CV
keywords kneeosteoarthritistotalreplacementpredictionMRIdomaingeneralizationinstancenormalizationcontrastivelearningimageaugmentationmedicalanalysis
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper claims that a deep-learning model predicting total knee replacement from knee MRI can be made to generalize across MRI protocols without any labeled target data. In the authors' setup, the baseline classifier, trained on fat-suppressed intermediate-weighted turbo spin-echo images and tested on dual-echo steady-state images, collapses to near-chance performance in the target domain (accuracy about 53%, F1 about 19%, ROC AUC about 59%). Replacing batch normalization with instance normalization, generating augmented views with global intensity non-linear augmentation, and adding a supervised contrastive loss raises target accuracy to about 70%, F1 to 67%, and ROC AUC to 78%, while also improving source-domain metrics. The paper reports that these gains are statistically significant and that the combination outperforms the evaluated single-source domain generalization methods under the same training setup.

What carries the argument

The load-bearing mechanism is the interaction of three training-time components. Instance normalization normalizes each 3D volume by its own statistics, removing the batch-dependent intensity statistics that differ between MRI protocols. GIN augmentation passes each volume through a shallow random-weight convolutional network with tanh activations and interpolates the result with the original volume using a mixing coefficient drawn from $U(0,1)$, creating intensity and texture variation that preserves anatomy. A supervised contrastive loss, $L = [d_p - m_{pos}]_+ + [m_{neg} - d_n]_+$, is applied to L2-normalized 256-dimensional feature vectors with margins $m_{pos}=0$ and $m_{neg}=1$, using four augmented views per sample plus same-label views from the mini-batch as positives and different-label views as negatives. The fourth mechanism is model selection: checkpoints must pass a source validation accuracy threshold (starting at 0.75 and lowered in steps of 0.05) and minimize prediction entropy on the unlabeled target validation set, an application of the low-density separation assumption.

What would settle it

Train the same three-component model on the same source and target MRI pair but select checkpoints by best source validation accuracy or by a fixed late epoch instead of by target-validation entropy, then measure target test accuracy; if it drops back toward the 50-60% range, the entropy criterion, not the representation learning, is carrying most of the reported gain. A complementary check is to keep the entropy rule but compare it against a random choice among checkpoints that pass the source accuracy threshold; if random selection performs about as well, the low-density separation assumption is not doing load-bearing work.

Watch

Extended reading notes

Core claim

The central claim is that three changes to a standard 3D convolutional classifier and its training loop make an MRI-based total knee replacement predictor robust to a change in imaging sequence: instance normalization in place of batch normalization, global intensity non-linear (GIN) augmentation that interpolates each input volume with the output of a fixed random-weight nonlinear network, and a supervised contrastive loss that pulls together augmented views and same-label samples while pushing apart different-label samples. In the target domain, the combined model reaches 70.04 ± 2.49 accuracy, 67.30 ± 3.57 F1, and 78.12 ± 1.97 ROC AUC, compared with 52.87 ± 3.17, 18.98 ± 16.89, and 59.33 ± 6.20 for the baseline. The source domain also improves on all three metrics. Ablation studies attribute the effect to the combination: without instance normalization, GIN plus contrastive loss reaches only 57.90 target accuracy, while the full model reaches 70.04. The authors also find that the contrastive loss with margins (0,1) outperforms InfoNCE, NCA, and triplet losses in this small-batch 3D setting.

Load-bearing premise

The load-bearing premise is that a model's average uncertainty on unlabeled target validation images is a reliable guide to its true target accuracy, and that the reused fold splits from the earlier study are faithful; if either assumption fails, the reported target gains could be inflated.

Editorial extensions

If this is right

  • A TKR classifier trained on one MRI sequence can be applied to another sequence without collecting new labels, because none of the three components uses target labels during training.
  • Target-domain F1 rises from roughly 19 to 67, meaning the model stops collapsing to the no-TKR class when the imaging protocol changes.
  • The combined method is reported to outperform the evaluated single-source domain generalization alternatives when instance normalization is used, so the components do not simply trade source performance for target performance.
  • Ablations show that instance normalization is the largest single contributor: removing it drops target accuracy from 70.04 to 57.90 even with GIN and contrastive loss still active.
  • The interpolation coefficient in GIN is a real tuning knob: $\alpha=1$ (extreme augmentation) gives the worst source accuracy, while $U(0,1)$ mixing gives the best target accuracy.

Reading between the lines

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

  • The same recipe should transfer to other MRI tasks that suffer from protocol shift, such as cartilage lesion segmentation or meniscal tear classification, since the contrastive loss only needs image-level labels.
  • The largest reported gain comes from fixing the baseline's target-domain collapse to the majority class; on a protocol shift that does not produce severe class imbalance, the method's advantage over a well-calibrated baseline would likely be smaller, and this is testable with the same ablation.
  • The model-selection rule uses unlabeled target-validation entropy, but the paper does not report what happens with simpler rules such as last-epoch selection or best source-validation accuracy; comparing those rules would separate the contribution of representation learning from the contribution of checkpoint selection.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 5 minor

Summary. The manuscript studies single-source domain generalization for MRI-based total knee replacement (TKR) prediction using the OAI dataset, with FS-IW-TSE as the source domain and DESS as the target domain. The proposed modifications are replacing batch normalization with instance normalization, adding GIN-based intensity augmentation with four augmented views per input, and adding a supervised contrastive loss to the classification loss. On the authors' seven-fold protocol, the full model improves target accuracy from 52.87 ± 3.17 to 70.04 ± 2.49, F1 from 18.98 ± 16.89 to 67.30 ± 3.57, and ROC AUC from 59.33 ± 6.20 to 78.12 ± 1.97, with smaller source-domain improvements. The paper includes ablations on contrastive-loss margins, metric-learning losses, and the GIN interpolation coefficient, and compares against RAC, MixStyle, CutOut, RSC, and AdvBias.

Significance. If the reported target-domain gains survive a target-blind evaluation protocol, the result is practically valuable: modest changes to normalization, augmentation, and the training objective substantially improve MRI-sequence transfer for a clinically relevant endpoint. The comparison includes several established domain-generalization methods, the target-domain improvements are consistent across folds, and the target accuracy, F1, and AUC improvements are statistically strong under paired tests. The main limitation is that model selection uses unlabeled target-domain validation entropy, which makes the current evaluation transductive rather than strictly single-source domain generalization; resolving this issue is essential before the headline claim can be accepted.

major comments (5)
  1. [IV (Model selection)] The model-selection protocol described in Section IV uses unlabeled target-domain validation entropy as the operative criterion. Because the source-accuracy threshold is 'initially set to 0.75; if this threshold was not met, it was progressively lowered by 0.05,' the threshold is always satisfied by construction, so the only constraining criterion is minimum entropy on the target validation set. This means the reported target test numbers in Table III are conditional on access to target-distribution data, which is an unsupervised domain adaptation/transductive setup rather than the single-source domain generalization claimed in the abstract. If the low-density separation assumption does not hold for this task, the selected epoch can be confidently wrong, and the improvement from 52.87% to 70.04% target accuracy may reflect target-informed selection rather than intrinsic generalization of IN+GIN+contrastive loss. Please re-evaluate with model selection based only on source validation accuracy (or a fixed epoch) and report whether the target gains persist; additionally, provide evidence for the low-density separation assumption, for example by correlating target validation entropy with target test accuracy across epochs.
  2. [Abstract and Table III] The abstract states that 'The GIN method with contrastive loss performed better than all evaluated single-source domain generalization methods when using 3D instance normalization.' Table III directly contradicts this for accuracy: AdvBias+IN achieves 70.26 ± 3.38 target accuracy while GIN+ContrastiveLoss+IN achieves 70.04 ± 2.49. The claimed superiority is therefore metric-dependent and should be qualified (e.g., 'higher F1 and AUC, comparable accuracy') or supported by a formal comparison with confidence intervals or paired tests.
  3. [V (Statistical significance)] The text reports one-sided paired t-tests and states that 'the performance improvements are statistically significant' in both domains. However, the source-domain ROC AUC p-value is 8.37 × 10⁻², which is not significant at α = 0.05. The abstract's claim of 'statistically significant improvement in classification metrics across both domains' is thus too strong; either report significance per metric accurately or adjust the claim.
  4. [II (Dataset) and IV (Method)] The evaluation protocol is not described precisely enough to be reproduced. The paper states that the seven-fold splits follow [9], where 'each fold consisted of six training splits, six validation splits, and one test set,' but it then designates 'the first validation split' as the target validation set and subdivides 'the first training split' into source training and source validation sets. It is unclear how the remaining five training and validation splits are used, how the 100-sample source validation set is drawn, and whether the same protocol is applied to all compared methods. Please provide a complete algorithmic description of the splits or release the exact fold indices, since the cross-validation results are the only evidence for the central claim.
  5. [V (Ablations, Tables V–VII)] The hyperparameters of the final model (m_pos = 0, m_neg = 1, ContrastiveLoss, α ~ U(0,1)) appear to have been selected after inspecting target-domain results in Tables V–VII. If any of these comparisons used the target test set, the reported test metrics and p-values are optimistically biased because the test set has effectively been used for model selection. Please clarify whether the ablation comparisons were made on the target validation set or the target test set, and if target test information was used, apply a nested validation scheme or report the selection procedure explicitly.
minor comments (5)
  1. [IV (Equation 2)] The text states that the L2 distance between normalized feature vectors lies in the range [0,1]; for unit-normalized vectors, the L2 distance lies in [0,2]. This should be corrected, although it does not change the main results.
  2. [Throughout] There are several typographical issues: 'ContrasitveLoss' in Section IV, 'ConstrastiveLoss' in Table V, 'differential blocks' should be 'residual blocks' in Section V, and the title contains a spurious space in 'T otal.'
  3. [II (Dataset)] The phrase 'no samples from the same domain in more than one domain' is confusing; please rephrase to state clearly that the source and target sets contain disjoint subjects or disjoint imaging sessions as intended.
  4. [V (Table III)] The abbreviation RAC is used in Table III but defined only later in Section V; please define it in the table caption or at first use.
  5. [IV (Loss function)] The description of positive pairs is slightly ambiguous: 'its four augmented versions and the four augmented versions of all other samples in the mini-batch that share the same label' could be read as including the anchor's own augmented versions twice if the anchor is also 'another sample.' Please clarify the pairing rule.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: empirical comparison built from external methods; target-entropy checkpoint selection is a protocol caveat, not a by-construction reduction.

full rationale

The claimed improvement is an empirical comparison of training modifications (instance normalization, GIN augmentation, supervised contrastive loss) that are each drawn from external prior work ([18]-[20], [23], [24]), not derived from the target-domain results. The baseline and proposed models are evaluated on test folds inherited from an external study ([9]), and no loss term, architecture component, or evaluation metric is defined in terms of the reported target accuracy, F1, or AUC. The only potentially circular-looking element is checkpoint selection via entropy on the target validation set in Section IV. However, target validation entropy is not equal by construction to target test accuracy or F1; it is a heuristic model-selection criterion, so this is an evaluation-protocol caveat that could inflate reported gains, not a circular reduction. The source-accuracy threshold is described as being progressively lowered if not met, which makes the entropy criterion operative, but again this is a selection-bias concern rather than a definitional equivalence. The hyperparameter ablations in Tables V-VII may involve selection among configurations, but the paper does not state that the final configuration was selected on the test set, and no equation-level equivalence is exhibited. No load-bearing self-citations by the present authors appear in the argument chain. Therefore no circular steps are established.

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

The paper contributes no new theory or entities; it assembles existing components. The central claim rests on empirical assumptions about normalization, augmentation, and target-validation-based model selection rather than on novel axioms.

free parameters (4)
  • Contrastive loss margins (m_pos, m_neg) = 0 and 1
    Table V ablation shows the (0,1) setting gives the best target accuracy and F1, so the reported main result uses the best of three tested margin pairs.
  • GIN interpolation coefficient alpha = sampled from U(0,1)
    Table VII ablation selects U(0,1) over alpha=1, Beta(0.5,0.5), and Beta(4,4) based on target performance.
  • Source validation accuracy threshold for model selection = 0.75, lowered by 0.05 increments
    Ad hoc rule used along with target entropy to pick the best model; the threshold is not justified by a principled criterion.
  • Number of GIN augmented views = 4
    Set by the authors without an ablation; directly affects the contrastive loss batch composition and training cost.
assumptions (4)
  • domain assumption Low-density separation assumption: the optimal decision boundary lies in low data-density regions, so target validation entropy is a valid model-selection criterion.
    Invoked for model selection in Section IV using the entropy criterion from [25]. If false, the selected model may not generalize to the target test set.
  • domain assumption MRI intensity variations can be decomposed into anatomical structure and domain-specific appearance, so intensity augmentation (GIN) preserves task-relevant structure.
    Stated in Section IV with refs [21,22]; underlies the use of GIN augmentation to improve robustness.
  • domain assumption Instance normalization removes domain-specific statistics and is therefore preferable to batch normalization under domain shift.
    Based on prior studies [18,19]; the paper does not demonstrate this specifically for TKR-relevant cartilage texture features.
  • domain assumption Source and target test sets, though derived from the same participants, are sufficiently independent to serve as valid domain-generalization test data.
    Section II treats FS-IW-TSE and DESS sequences from the same knee as different domains. If the two sequences share subject-specific confounders, the measured gain may be optimistic.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Improving Generalization in MRI-Based Deep Learning Models for Total Knee Replacement Prediction." pith.science (2026). https://pith.science/paper/MW7OR37Z

@misc{pith2026250419203,
  author       = {Pith},
  title        = {Pith review of: Improving Generalization in MRI-Based Deep Learning Models for Total Knee Replacement Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MW7OR37Z}},
  note         = {Machine review of arXiv:2504.19203}
}
read the original abstract

Knee osteoarthritis (KOA) is a common joint disease that causes pain and mobility issues. While MRI-based deep learning models have demonstrated superior performance in predicting total knee replacement (TKR) and disease progression, their generalizability remains challenging, particularly when applied to imaging data from different sources. In this study, we show that replacing batch normalization with instance normalization, using data augmentation, and applying contrastive loss improves generalization. For training and evaluation, we used MRI data from the Osteoarthritis Initiative (OAI) database, considering sagittal fat-suppressed intermediate-weighted turbo spin-echo (FS-IW-TSE) images as the source domain and sagittal fat-suppressed three-dimensional (3D) dual-echo in steady state (DESS) images as the target domain. The results demonstrated a statistically significant improvement in classification metrics across both domains by replacing batch normalization with instance normalization in the baseline model, generating augmented input views using the Global Intensity Non-linear (GIN) augmentation method, and incorporating a supervised contrastive loss alongside the classification loss to align representations of samples with the same label. The GIN method with contrastive loss performed better than all evaluated single-source domain generalization methods when using 3D instance normalization. Comparing GIN with and without contrastive loss (for both normalization types) showed that adding contrastive loss consistently led to better performance.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 17 canonical work pages

  1. [9]

    Prediction of total knee replacement using deep learning analysis of knee MRI,

    H. R. Rajamohan et al., “Prediction of total knee replacement using deep learning analysis of knee MRI,” Scientific Reports, vol. 13, no. 1, Apr. 2023, doi: 10.1038/s41598-023-33934-1

    work page doi:10.1038/s41598-023-33934-1 2023

  2. [1]

    Deep learning for knee osteoarthritis diagnosis and progression prediction from plain radiographs and clinical data,

    A. Tiulpin, “Deep learning for knee osteoarthritis diagnosis and progression prediction from plain radiographs and clinical data,”

    work page

  3. [2]

    Segmentation of knee bones for Osteoarthritis assessment: A comparative analysis of supervis ed, Few -Shot, and Zero -Shot learning approaches,

    Y. X. Teoh, A. Othmani, S. L. Goh, J. Usman, and K. W. Lai, “Segmentation of knee bones for Osteoarthritis assessment: A comparative analysis of supervis ed, Few -Shot, and Zero -Shot learning approaches,” in IFMBE proceedings, 2024, pp. 331 –339. doi: 10.1007/978-3-031-62523-7_37

    work page doi:10.1007/978-3-031-62523-7_37 2024

  4. [3]

    A Progressive Risk Formulation for Enhanced Deep Learning based Total Knee Replacement Prediction in Knee Osteoarthritis

    H. R. Rajamohan, R. Kijowski, K. Cho, and C. M. Deniz, “Modified risk formulation for improving the prediction of knee osteoarthr itis progression,” arXiv, Jun. 2024, doi: 10.48550/arxiv.2406.10119

    work page Pith review arXiv doi:10.48550/arxiv.2406.10119 2024

  5. [4]

    Discovering knee osteoarthritis imaging features for diagnosis and prognosis: Review of manual imaging grading and machine learning approaches,

    Y. X. Teoh et al., “Discovering knee osteoarthritis imaging features for diagnosis and prognosis: Review of manual imaging grading and machine learning approaches,” Journal of Healthcare Engineering, vol. 2022, pp. 1–19, Feb. 2022, doi: 10.1155/2022/4138666

    work page doi:10.1155/2022/4138666 2022

  6. [5]

    Imaging of knee osteoarthritis: a review of multimodal diagnostic appr oach,

    Claudia Lucia Piccolo, Carlo Augusto Mallio, F. Vaccarino, Rosario Francesco Grasso, and Bruno Beomonte Zobel, “Imaging of knee osteoarthritis: a review of multimodal diagnostic appr oach,” Quantitative imaging in medicine and surgery, vol. 13, no. 11, pp. 7582–7595, Nov. 2023, doi: https://doi.org/10.21037/qims-22-1392

    work page doi:10.21037/qims-22-1392 2023

  7. [6]

    A deep learning method for predicting knee osteoarthritis radiographic progression from MR I,

    J.-B. Schiratti et al., “A deep learning method for predicting knee osteoarthritis radiographic progression from MR I,” Arthritis Research & Therapy, vol. 23, no. 1, Oct. 2021, doi: 10.1186/s13075- 021-02634-4

    work page doi:10.1186/s13075- 2021

  8. [7]

    Deep Learning Predicts Total Knee Replacement from Magnetic Resonance Images,

    A. A. Tolpadi, J. J. Lee, V. Pedoia, and S. Majumdar, “Deep Learning Predicts Total Knee Replacement from Magnetic Resonance Images,” Scientific Reports, vol. 10 , no. 1, Apr. 2020, doi: 10.1038/s41598-020-63395-9

    work page doi:10.1038/s41598-020-63395-9 2020

Show all 29 references
  1. [8]

    Predicting Knee Osteoarthritis Progression from Structural MRI Using Deep Learning,

    E. Panfilov, S. Saarakkala, M. T. Nieminen, and A. Tiulpin, “Predicting Knee Osteoarthritis Progression from Structural MRI Using Deep Learning,” 2022 IEEE 19th International Symposium on Biomedical Ima ging (ISBI), vol. 9, pp. 1 –5, Mar. 2022, doi: 10.1109/isbi52829.2022.9761458

    2022

  2. [10]

    Unsupervised domain adaptation for automated knee osteoarthritis phenotype classification,

    J. Zhong et al., “Unsupervised domain adaptation for automated knee osteoarthritis phenotype classification,” Quantitative Imaging in Medicine and Surgery, vol. 13, no. 11, pp. 7444–7458, Oct. 2023, doi: https://doi.org/10.21037/qims-23-704

    2023 doi

  3. [11]

    Unsupervised Domain Adaptation by Backpropagation,

    Y. Ganin and V. Lempitsky, “Unsupervised Domain Adaptation by Backpropagation,” arXiv.org, Feb. 27, 2015. https://arxiv.org/abs/1409.7495

    2015 arXiv

  4. [12]

    Deep Domain Confusion: Maximizing for Domain Invariance,

    E. Tzeng, J. Hoffman, N . Zhang, K. Saenko, and T. Darrell, “Deep Domain Confusion: Maximizing for Domain Invariance,” arXiv.org, Dec. 10, 2014. https://arxiv.org/abs/1412.3474 (accessed Sep. 09, 2023)

    2014 arXiv

  5. [13]

    Domain Generalization via Invariant Feature Representation,

    Krikamol Muandet, D. Balduzzi, and Bernhard Sch lkopf, “Domain Generalization via Invariant Feature Representation,” arXiv (Cornell University), Jan. 2013, doi: https://doi.org/10.48550/arxiv.1301.2115

    2013 doi arXiv

  6. [14]

    Improved Regularization of Convolutional Neural Networks with Cutout,

    T. Devries and G. W. Taylor, “Improved Regularization of Convolutional Neural Networks with Cutout,” arXiv (Cornell University), Jan. 2017, doi: https://doi.org/10.48550/arxiv.1708.04552

    2017 doi arXiv

  7. [15]

    MixStyle Neural Networks for Domain Generalization and Adaptation,

    K. Zhou, Y. Yang, Y. Qiao, and T. Xiang, “MixStyle Neural Networks for Domain Generalization and Adaptation,” arXiv (Cornell University), J an. 2021, doi: https://doi.org/10.48550/arxiv.2107.02053

    2021 doi arXiv

  8. [16]

    Realistic Adversarial Data Augmentation for MR Image Segmentation,

    C. Chen et al., “Realistic Adversarial Data Augmentation for MR Image Segmentation,” arXiv (Cornell University), Jan. 2020, doi: https://doi.org/10.48550/arxiv.2006.13322

    2020 doi arXiv

  9. [17]

    Self -Challenging Improves Cross -Domain Generalization,

    Z. Huang, H. Wang, E. P. Xing, and D. Huang, “Self -Challenging Improves Cross -Domain Generalization,” arXiv (Cornell University), Jul. 2020, doi: https://doi.org/10.48550/arxiv.2007.02454

    2020 doi arXiv

  10. [18]

    Learning to optimize domain specific normalization for domain generalization,

    S. Seo, Y. Suh, D. Kim, G. Kim, J. Han, and B. Han, “Learning to optimize domain specific normalization for domain generalization,” in Lecture notes in computer science, 2020, pp. 68 –83. doi: 10.1007/978-3-030-58542-6_5

    2020 doi

  11. [19]

    Adversarially adaptive normalization for single domain generalization,

    X. Fan, Q. Wang, J. Ke, F. Yang, B. Gong, and M. Zhou, “Adversarially adaptive normalization for single domain generalization,” arXiv, Jan. 2021, doi: 10.48550/arxiv.2106.01899

    2021 doi arXiv

  12. [20]

    Causality -Inspired Single -Source domain generalization for medical image segmentation,

    C. Ouyang et al., “Causality -Inspired Single -Source domain generalization for medical image segmentation,” IEEE Transactions on Medical Imaging, vol. 42, no. 4, pp. 1095–1106, Nov. 2022, doi: 10.1109/tmi.2022.3224067

    2022

  13. [21]

    Unsupervised MR harmonization by learning disentangled representations using information bottleneck theory,

    L. Zuo et al., “Unsupervised MR harmonization by learning disentangled representations using information bottleneck theory,” NeuroImage, vol. 243, p. 118569, Sep. 2021, doi : 10.1016/j.neuroimage.2021.118569

    2021

  14. [22]

    Disentangling a single MR modality,

    L. Zuo et al., “Disentangling a single MR modality,” in Lecture notes in computer science, 2022, pp. 54 –63. doi: 10.1007/978 -3-031- 17027-0_6

    2022 doi

  15. [23]

    PyTorch Metric Learning,

    K. Musgrave, S. Belongie, and S. -N. Lim, “PyTorch Metric Learning,” arXiv, Jan. 2020, doi: 10.48550/arxiv.2008.09164

    2020 doi arXiv

  16. [24]

    Adversarial Consistency for Single Domain Generalization in Medical Image Segmentation,

    Y. Xu, S. Xie, M. Reynolds, M. Ragoza, M. Gong, and K. Batmanghelich, “Adversarial Consistency for Single Domain Generalization in Medical Image Segmentation,” Lecture Notes in Computer Science, pp. 671 –681, 2022, doi: https://doi.org/10.1007/978-3-031-16449-1_64

    2022 doi

  17. [25]

    Mixed Samples as Probes for Unsupervised Model Selection in Domain Adaptation,

    D. Hu, J. Liang, J. H. Liew, C. Xue, S. Bai, and X. Wang, “Mixed Samples as Probes for Unsupervised Model Selection in Domain Adaptation,” Advances in Neural Information Processing Syste ms, vol. 36, pp. 37923 –37941, 2023. https://proceedings.neurips.cc/paper_files/paper/2023...

    2023

  18. [26]

    Minimal-Entropy Correlation Alignment for Unsupervised Deep Domain Adaptation,

    Pietro Morerio, Jacopo Cavazza, and V. Murino, “Minimal-Entropy Correlation Alignment for Unsupervised Deep Domain Adaptation,” International Conference on Learning Representations (ICLR) 2018, Apr. 2018, doi: https://doi.org/10.48550/arXiv.1711.10288

    2018 doi arXiv

  19. [27]

    Three New Validators and a Large -Scale Benchmark Ranking for Unsuper vised Domain Adaptation,

    K. Musgrave, S. Belongie, and S. -N. Lim, “Three New Validators and a Large -Scale Benchmark Ranking for Unsuper vised Domain Adaptation,” arXiv (Cornell University), Jan. 2022, doi: https://doi.org/10.48550/arxiv.2208.07360

    2022 doi arXiv

  20. [28]

    Assessing Model Out- of-distribution Generalization with Softmax Prediction Probability Baselines and A Correlation Method,

    W. Tu, W. Deng, T. Gedeon, and L. Zheng, “Assessing Model Out- of-distribution Generalization with Softmax Prediction Probability Baselines and A Correlation Method,” Openreview.net, 2023. https://openreview.net/forum?id=1maXoEyeqx

    2023

  21. [2020]

    https://api.semanticscholar.org/CorpusID:219128608

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.