Pith. sign in

REVIEW 3 major objections 4 minor 42 references

Glioma Multimodal MRI Analysis System for Tumor Layered Diagnosis via Multi-task Semi-supervised Learning

T0 review · 3 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read The paper argues that a single multi-task network can perform the layered diagnosis of gliomas from multimodal MRI, simultaneously segmenting tumor regions and predicting histological subtype, IDH mutation status, and 1p/19q status, with…

desk verdict The integration is plausible and the ablations are clean, but the unstated BraTS-to-TCGA label linkage makes the headline AUCs unverifiable until the authors show their data provenance. read the letter →

arxiv 2501.17758 v2 pith:PLDUMIWD submitted 2025-01-29 eess.IV cs.CV

Yihao Liu , Zhihao Cui , Liming Li , Junjie You , Xinle Feng , Jianxin Wang , Xiangyu Wang , Qing Liu
show 1 more author
Minghua Wu
This is my paper · ORCID
classification eess.IVcs.CV
keywords gliomamultimodalMRImulti-tasklearningsemi-supervisedtumorsegmentationIDHmutation1p/19qcodeletionmodalityabsence
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 tries to establish that glioma diagnosis from MRI should be treated as one layered problem, not a stack of separate tools. Its system, GMMAS, claims to segment the tumor and predict the histological subtype (GBM vs LGG), IDH mutation status, 1p/19q codeletion status, and MGMT status in a single forward pass, using a shared feature representation. The authors report that this multi-task design, together with a two-stage semi-supervised learning scheme, beats their single-task baselines and reaches AUC values of 0.980 for subtype, 0.982 for IDH, and 0.944 for 1p/19q. They also claim an adaptation module makes the system degrade gracefully when one or two MRI modalities are missing. If those numbers hold up under clean label linkage, the clinical payoff is one MRI-based workflow that gives surgeons and oncologists the layered molecular picture currently spread over several separate analyses.

What carries the argument

The load-bearing object is a multi-task CNN-transformer encoder with two uncertainty mechanisms. Aleatoric uncertainty, as trainable loss weights, balances the segmentation and classification tasks; epistemic uncertainty, estimated by MC-Dropout, filters which unlabeled scans receive pseudo-labels. Around this core sit a global-feature U-net fused by channel attention, a Tumor-CutMix augmentation that mixes tumor patches and soft labels to improve calibration, a two-stage semi-supervised loop (dual-threshold pseudo-labeling plus weak-to-strong consistency regularization), and an adaptation module that trains a full-modality 'teacher' pathway and a missing-modality 'student' pathway against each other.

What would settle it

Re-run the classification heads on an external cohort whose IDH and 1p/19q status was confirmed by tissue sequencing and linked to MRI by patient identifier; if AUCs fall near 0.5 while segmentation Dice stays high, the reported classification precision came from label leakage or dataset overlap rather than from MRI features.

Watch

Extended reading notes

Core claim

GMMAS's central claim is that tumor segmentation and molecular subtyping are mutually reinforcing, so training them jointly in one network yields higher accuracy than dedicated single-task models. The paper reports whole-tumor Dice of 0.940 ± 0.037 and classification AUCs of 0.980 for histological subtype, 0.982 for IDH, and 0.944 for 1p/19q after adding global feature fusion and two-stage semi-supervised learning. It also argues that a cross-modal adaptation module based on knowledge self-distillation and contrastive learning lets the model maintain useful segmentation and classification performance even when individual MRI sequences are absent, and that the learned uncertainty values rank tasks by difficulty, with IDH and grade easier than MGMT and 1p/19q.

Load-bearing premise

The system assumes that every MRI volume is correctly paired with its molecular and histological ground truth, so if the imaging-to-label matching is wrong or the two releases of the public challenge dataset share patients between training and test, every reported AUC and the external validation are invalid.

Editorial extensions

If this is right

  • One MRI analysis system could replace the current sequence of separate segmentation, grading, and genotyping pipelines.
  • The reported AUC values imply IDH status and glioma grade are readable from MRI morphology at clinically relevant accuracy, while MGMT and 1p/19q remain harder.
  • The modality-absence results imply a trained model could still produce useful diagnoses when a hospital lacks a particular MRI sequence, such as T2-FLAIR or T1CE.
  • Two-stage semi-supervised learning turns the large pool of scans without molecular labels into training signal, easing the annotation bottleneck that limits radiogenomic models.
  • Calibration via Tumor-CutMix means the confidence shown to clinicians alongside each prediction is more trustworthy, not just the prediction itself.

Reading between the lines

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

  • If the claimed label linkage is clean, the same multi-task architecture should transfer to other central-nervous-system tumors such as meningiomas and ependymomas, where the same layered diagnostic logic applies.
  • The modality-importance ranking the paper reports (T2-FLAIR for edema, T1CE for tumor core) suggests MRI acquisition protocols could be shortened or prioritized for these sequences without sacrificing diagnosis.
  • The correlation the paper draws between epistemic uncertainty and calibration error could be tested as a general model-selection rule: hold back predictions whose MC-Dropout standard deviation exceeds a threshold.
  • A direct head-to-head against a modern single-task foundation model on the same held-out cohort would be the natural next check of whether joint training is what buys the accuracy gain.
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 / 4 minor

Summary. The paper presents GMMAS, a multi-task deep learning system that jointly performs glioma segmentation, histological subtype classification, and prediction of IDH, 1p/19q, and MGMT molecular markers from multimodal MRI. The method combines a CNN-transformer encoder, a U-Net segmentation decoder, a multi-task classifier, uncertainty-based task weighting, a global feature fusion module, and a two-stage semi-supervised learning strategy with pseudo-labeling and consistency regularization. An adaptation module based on knowledge self-distillation and contrastive learning is introduced to handle missing MRI modalities, and a clinical platform with a GPT-based prognostic report generator is described. The central claims are that GMMAS improves precision over reported single-task models across layered diagnostic tasks and that it remains robust when modalities are absent.

Significance. If the results hold, the system addresses a genuinely clinically relevant problem: integrating segmentation, grading, and radiogenomic classification into one platform with graceful degradation under missing modalities. The internal ablations (Tables II, III, A6, A7) are internally consistent and show monotonic gains as modules are added, and the move toward multi-task and semi-supervised learning for glioma MRI is timely. The paper also provides a fairly detailed description of the architecture and training procedure, including explicit loss functions (Eqs. 3–11), which facilitates reproduction. However, the central claim of superiority over reported single-task models is not supported by any external comparison, and the molecular label linkage to imaging data is not described, leaving the main classification results unverifiable as currently presented.

major comments (3)
  1. [III-A, Table I] The paper states that molecular marker information for BraTS patients was 'obtained from The Cancer Genome Atlas (TCGA) gene database,' and that the TCIA collection LGG-1p19qDeletion was also used, but it never describes the patient-level linkage between the MRI imaging files and the TCGA molecular records. It also does not specify how overlapping patients between BraTS 2020 and BraTS 2021 were handled or whether any overlap created train/test leakage. Because every classification AUC in Table III and the ROCs in Figure 10 depend on these labels being correctly assigned to the corresponding MRI volumes, the lack of this information is a load-bearing gap. Please provide the linkage procedure, the exact sources of each label, and the overlap-removal strategy, or the classification results cannot be considered verified.
  2. [Abstract and Section IV-B-2] The abstract claims that 'Compared with the reported single-task analysis models, GMMAS improves the precision across tumor layered diagnostic tasks,' but the paper does not compare against any published single-task method or external benchmark numbers. Table A6 compares only against the authors' own single-task baseline, which is not a 'reported' model in the literature. Please either add comparisons with published results on the same or comparable datasets (with proper handling of dataset differences) or substantially soften the claim so that it is limited to the internal ablation.
  3. [Tables III and I] The classification results in Table III are reported as point estimates without confidence intervals or significance tests, and the number of test subjects with valid labels for each task is not reported. Given the small label cohorts in Table I (e.g., only 71 IDH-mutant training samples and 21 IDH-mutant internal validation samples), the improvements from +global feature fusion to +semi-supervised learning could be within sampling noise. Please report confidence intervals or at least the per-task test sample sizes, and consider statistical testing for the reported AUC differences.
minor comments (4)
  1. [Throughout] The title and several headings contain typos: 'Tumor' is rendered as 'T umor' in the title, and 'modals' appears where 'modalities' is meant (e.g., in the caption of Figure 4). Please proofread the text.
  2. [Eq. (1)] The modality weight notation contains a typo: 'Wf lair' should be 'Wflair', and the subscript for T1CE appears inconsistently as 'T lce' in one place. Please standardize the subscripts.
  3. [Fig. A12 caption] The caption references 'Murray et al., 2014' for global average pooling and global max pooling, but this reference is not included in the reference list. Please add the citation or remove the attribution.
  4. [Algorithm 1] The pseudocode for the tumor filter is difficult to interpret; for example, the variable 'Threshold' is used both as a boolean condition and as a comparison value inside the while loop. Please rewrite the algorithm with clear variable names and unambiguous control flow.

Circularity Check

0 steps flagged · score 1.0 of 10

Empirical multi-task system; no Eq. X = Eq. Y circularity; self-training loop and label-linkage concerns are validity issues, not circularity.

full rationale

This is an empirical systems paper, not an analytic derivation, so the circularity patterns based on equations reducing to their own inputs do not apply. The claimed improvements (multi-task over single-task, global feature fusion, two-stage semi-supervised learning, and modality-absence adaptation) are supported by held-out quantitative comparisons in Tables II, III, A6, A7, and Figure 10, not by equations that replay the training inputs. The loss functions in Eqs. 3-11 are standard training objectives with learned uncertainty weights; no reported AUC or Dice score is algebraically forced by a fitted parameter. The semi-supervised stages do use the model's own confident predictions as pseudo-labels in the first stage and EMA-consistency targets in the second (Figure 6), which is self-referential during training, but the final evaluation is performed on held-out labels, so the stated results are not circular by construction. Author self-citations such as [4], [6], [7], and [14] appear in background and related-work statements and are not load-bearing for the central claim. The weakest point is the data-assembly step in Section III-A: molecular marker information is said to come from TCGA/TCIA without describing the patient-level linkage or duplicate handling between BraTS 2020 and 2021, which threatens the validity of the classification results. This is a data-integrity and correctness concern, not a circularity concern under the rubric. No step was found where a prediction is definitionally equivalent to its input, so the score is 1 rather than 0 only to acknowledge the self-training loop and the unverified label-linkage assumption.

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

The central empirical claims rest mostly on data-quality assumptions (label linkage, MRI-molecular correlation), on three hand-selected thresholds (pseudo-label thresholds, tumor filter), and on learned weighting parameters. None of these are derived from first principles; they are standard practice in applied deep learning but they limit how much the paper can claim beyond its own dataset.

free parameters (5)
  • Uncertainty thresholds for pseudo-labels = tau_u <= 0.1, tau_c >= 0.95
    Set in Section III-F to select reliable pseudo-labels for self-training; values are hand-chosen on the validation set and not derived from theory.
  • Tumor filter thresholds = D < 0.1; ratio < 0.6
    Algorithm 1 uses these hand-set thresholds to remove false positive tumor voxels; chosen heuristically, not optimized with a stated criterion.
  • Task uncertainty weights sigma = learned
    Equations 8-9 introduce trainable aleatoric uncertainty parameters that weight the multi-task loss; these are fitted to the training data.
  • Modality fusion weights W in Eq. 1 = learned
    The weighted average of modality feature maps uses learnable weights, described in Section III-C; these are fitted during training and are used to claim robustness to modality absence.
  • MC-Dropout pass count M = not reported
    Equation 6 defines epistemic uncertainty via M stochastic forward passes, but M is never specified; a free implementation choice affecting the pseudo-label filtering.
assumptions (5)
  • domain assumption Molecular and histological labels are correctly linked to the MRI volumes
    Section III-A says molecular markers were obtained from TCGA for BraTS patients and that the TCIA collection is LGG-1p19qDeletion, but no patient-level matching procedure is described. All classification results depend on this.
  • domain assumption Tumor molecular status is reflected in MRI morphology
    Stated in the Introduction: 'the molecular status can be characterized by MRI morphological features' and supported by cited literature; the entire classification approach relies on this.
  • ad hoc to paper High-confidence, low-uncertainty predictions on unlabeled data are accurate enough to serve as pseudo-labels
    Section III-F assumes predictions passing thresholds (tau_u, tau_c) are reliable; this is a modeling assumption specific to this training scheme.
  • domain assumption Combining BraTS 2020, BraTS 2021, TCIA, and Xiangya data does not introduce harmful distribution shift or patient overlap
    Section III-A combines four sources without discussing overlap (BraTS 2020 and 2021 share cases) or batch effects; external validity claims depend on this.
  • standard math Dice loss, weighted cross-entropy, and uncertainty-weighted loss are appropriate and correctly implemented
    Standard loss functions used in Eqs. 3-9, assumed correct.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Glioma Multimodal MRI Analysis System for Tumor Layered Diagnosis via Multi-task Semi-supervised Learning." pith.science (2026). https://pith.science/paper/PLDUMIWD

@misc{pith2026250117758,
  author       = {Pith},
  title        = {Pith review of: Glioma Multimodal MRI Analysis System for Tumor Layered Diagnosis via Multi-task Semi-supervised Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PLDUMIWD}},
  note         = {Machine review of arXiv:2501.17758}
}
read the original abstract

Gliomas are the most common primary tumors of the central nervous system. Multimodal MRI is widely used for the preliminary screening of gliomas and plays a crucial role in auxiliary diagnosis, therapeutic efficacy, and prognostic evaluation. Currently, the computer-aided diagnostic studies of gliomas using MRI have focused on independent analysis events such as tumor segmentation, grading, and radiogenomic classification, without studying inter-dependencies among these events. In this study, we propose a Glioma Multimodal MRI Analysis System (GMMAS) that utilizes a deep learning network for processing multiple events simultaneously, leveraging their inter-dependencies through an uncertainty-based multi-task learning architecture and synchronously outputting tumor region segmentation, glioma histological subtype, IDH mutation genotype, and 1p/19q chromosome disorder status. Compared with the reported single-task analysis models, GMMAS improves the precision across tumor layered diagnostic tasks. Additionally, we have employed a two-stage semi-supervised learning method, enhancing model performance by fully exploiting both labeled and unlabeled MRI samples. Further, by utilizing an adaptation module based on knowledge self-distillation and contrastive learning for cross-modal feature extraction, GMMAS exhibited robustness in situations of modality absence and revealed the differing significance of each MRI modal. Finally, based on the analysis outputs of the GMMAS, we created a visual and user-friendly platform for doctors and patients, introducing GMMAS-GPT to generate personalized prognosis evaluations and suggestions.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

42 extracted references · 31 canonical work pages

  1. [1]

    The 2016 world health organization classification of tu- mors of the central nervous system: a summary,

    D. N. Louis, A. Perry, G. Reifenberger, A. V on Deimling, D. Figarella- Branger, W. K. Cavenee, H. Ohgaki, O. D. Wiestler, P. Kleihues, and D. W. Ellison, “The 2016 world health organization classification of tu- mors of the central nervous system: a summary,” Acta neuropathologica, vol. 131, pp. 803–820, 2016

    work page 2016

  2. [2]

    The 2021 who classification of tumors of the central nervous system: a summary,

    D. N. Louis, A. Perry, P. Wesseling, D. J. Brat, I. A. Cree, D. Figarella- Branger, C. Hawkins, H. Ng, S. M. Pfister, G. Reifenberger et al. , “The 2021 who classification of tumors of the central nervous system: a summary,” Neuro-oncology, vol. 23, no. 8, pp. 1231–1251, 2021

    work page 2021

  3. [3]

    Glioma targeted therapy: insight into future of molecular approaches,

    K. Yang, Z. Wu, H. Zhang, N. Zhang, W. Wu, Z. Wang, Z. Dai, X. Zhang, L. Zhang, Y . Peng et al. , “Glioma targeted therapy: insight into future of molecular approaches,” Molecular Cancer, vol. 21, no. 1, p. 39, 2022

    work page 2022

  4. [4]

    Insight into deep learning for glioma idh medical image analysis: A systematic review,

    Q. Lv, Y . Liu, Y . Sun, and M. Wu, “Insight into deep learning for glioma idh medical image analysis: A systematic review,” Medicine, vol. 103, no. 7, p. e37150, 2024

    work page 2024

  5. [5]

    The implications of idh mutations for cancer development and therapy,

    C. J. Pirozzi and H. Yan, “The implications of idh mutations for cancer development and therapy,” Nature reviews Clinical oncology , vol. 18, no. 10, pp. 645–661, 2021

    work page 2021

  6. [6]

    Predictive panel for immunotherapy in low-grade glioma,

    Q. Lv, Z. Zhang, H. Fu, D. Li, Y . Liu, Y . Sun, and M. Wu, “Predictive panel for immunotherapy in low-grade glioma,” World Neurosurgery, vol. 183, pp. e825–e837, 2024

    work page 2024

  7. [7]

    A survey of embodied ai in healthcare: Techniques, applications, and opportunities,

    Y . Liu, X. Cao, T. Chen, Y . Jiang, J. You, M. Wu, X. Wang, M. Feng, Y . Jin, and J. Chen, “A survey of embodied ai in healthcare: Techniques, applications, and opportunities,” arXiv preprint arXiv:2501.07468, 2025

    arXiv 2025

  8. [8]

    Prediction of idh1-mutation and 1p/19q-codeletion status using preoperative mr imaging phenotypes in lower grade gliomas,

    Y . Park, K. Han, S. Ahn, S. Bae, Y . Choi, J. Chang, S. Kim, S.-G. Kang, and S.-K. Lee, “Prediction of idh1-mutation and 1p/19q-codeletion status using preoperative mr imaging phenotypes in lower grade gliomas,” American Journal of Neuroradiology , vol. 39, no. 1, pp. 37–42, 2018

    work page 2018

Show all 42 references
  1. [9]

    A fully automated multi- modal mri-based multi-task learning for glioma segmentation and idh genotyping,

    J. Cheng, J. Liu, H. Kuang, and J. Wang, “A fully automated multi- modal mri-based multi-task learning for glioma segmentation and idh genotyping,” IEEE Transactions on Medical Imaging, vol. 41, no. 6, pp. 1520–1532, 2022

    2022

  2. [10]

    Predicting pathological complete response based on weakly and semi-supervised joint learning from breast cancer mri,

    X. Hao, H. Xu, N. Zhao, T. Yu, T. Hamalainen, and F. Cong, “Predicting pathological complete response based on weakly and semi-supervised joint learning from breast cancer mri,” in2023 45th Annual International Conference of the IEEE Engineering in Medicine & Biology Society (...

    2023

  3. [11]

    Generalizable fgfr prediction across tumor types using self supervised learning

    C. Parmar, A. J. Ramon, N. L. Stone, S. Triantos, J. Greshock, and K. Standish, “Generalizable fgfr prediction across tumor types using self supervised learning.” 2023

    2023

  4. [12]

    Exploring simple siamese representation learning,

    X. Chen and K. He, “Exploring simple siamese representation learning,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 15 750–15 758

    2021

  5. [13]

    Multi-task learning using uncer- tainty to weigh losses for scene geometry and semantics,

    A. Kendall, Y . Gal, and R. Cipolla, “Multi-task learning using uncer- tainty to weigh losses for scene geometry and semantics,” inProceedings of the IEEE conference on computer vision and pattern recognition , 2018, pp. 7482–7491

    2018

  6. [14]

    Deep learning in precision medicine and focus on glioma,

    Y . Liu and M. Wu, “Deep learning in precision medicine and focus on glioma,” Bioengineering & Translational Medicine , vol. 8, no. 5, p. e10553, 2023

    2023

  7. [15]

    Convolutional neural networks: an overview and application in radiology,

    R. Yamashita, M. Nishio, R. K. G. Do, and K. Togashi, “Convolutional neural networks: an overview and application in radiology,” Insights into imaging, vol. 9, pp. 611–629, 2018

    2018

  8. [16]

    Deep learning in radiology: An overview of the concepts and a survey of the state of the art with focus on mri,

    M. A. Mazurowski, M. Buda, A. Saha, and M. R. Bashir, “Deep learning in radiology: An overview of the concepts and a survey of the state of the art with focus on mri,” Journal of magnetic resonance imaging , vol. 49, no. 4, pp. 939–954, 2019

    2019

  9. [17]

    Radiomics-based machine learning methods for isocitrate dehydrogenase genotype prediction of diffuse gliomas,

    S. Wu, J. Meng, Q. Yu, P. Li, and S. Fu, “Radiomics-based machine learning methods for isocitrate dehydrogenase genotype prediction of diffuse gliomas,” Journal of cancer research and clinical oncology , vol. 145, pp. 543–550, 2019

    2019

  10. [18]

    Noninvasive prediction of idh1 mutation and atrx expression loss in low-grade gliomas using multiparametric mr radiomic features,

    Y . Ren, X. Zhang, W. Rui, H. Pang, T. Qiu, J. Wang, Q. Xie, T. Jin, H. Zhang, H. Chen et al. , “Noninvasive prediction of idh1 mutation and atrx expression loss in low-grade gliomas using multiparametric mr radiomic features,” Journal of Magnetic Resonance Imaging , vol. 49, ...

    2019

  11. [19]

    Efficient multilevel brain tumor segmentation with integrated bayesian model classification,

    J. J. Corso, E. Sharon, S. Dube, S. El-Saden, U. Sinha, and A. Yuille, “Efficient multilevel brain tumor segmentation with integrated bayesian model classification,” IEEE transactions on medical imaging , vol. 27, no. 5, pp. 629–640, 2008

    2008

  12. [20]

    Radiomic feature clusters and prognostic signatures specific for lung and head & neck cancer,

    C. Parmar, R. T. Leijenaar, P. Grossmann, E. Rios Velazquez, J. Bussink, D. Rietveld, M. M. Rietbergen, B. Haibe-Kains, P. Lambin, and H. J. Aerts, “Radiomic feature clusters and prognostic signatures specific for lung and head & neck cancer,” Scientific reports, vol. 5, no. 1...

    2015

  13. [21]

    A robust classification of brain tumor disease in mri using twin-attention based dense convolutional auto- encoder,

    S. S. Waghere and J. P. Shinde, “A robust classification of brain tumor disease in mri using twin-attention based dense convolutional auto- encoder,” Biomedical Signal Processing and Control, vol. 92, p. 106088, 2024

    2024

  14. [22]

    Multi-task parallel with feature sharing integrated 3d u-nets for glioma segmentation,

    S. Li and S. Guo, “Multi-task parallel with feature sharing integrated 3d u-nets for glioma segmentation,” Biomedical Signal Processing and Control, vol. 93, p. 106178, 2024

    2024

  15. [23]

    Prediction of lower-grade glioma molecular subtypes using deep learning,

    Y . Matsui, T. Maruyama, M. Nitta, T. Saito, S. Tsuzuki, M. Tamura, K. Kusuda, Y . Fukuya, H. Asano, T. Kawamata et al. , “Prediction of lower-grade glioma molecular subtypes using deep learning,” Journal of neuro-oncology, vol. 146, pp. 321–327, 2020

    2020

  16. [24]

    Deep learning of imaging phenotype and genotype for predicting overall survival time of glioblastoma patients,

    Z. Tang, Y . Xu, L. Jin, A. Aibaidula, J. Lu, Z. Jiao, J. Wu, H. Zhang, and D. Shen, “Deep learning of imaging phenotype and genotype for predicting overall survival time of glioblastoma patients,” IEEE transactions on medical imaging , vol. 39, no. 6, pp. 2100–2109, 2020

    2020

  17. [25]

    Attention is all you need,

    A. Vaswani, “Attention is all you need,” Advances in Neural Information Processing Systems, 2017

    2017

  18. [26]

    Automatic multi-anatomical skull structure segmentation of cone-beam computed tomography scans using 3d unetr,

    M. Gillot, B. Baquero, C. Le, R. Deleat-Besson, J. Bianchi, A. Ruellas, M. Gurgel, M. Yatabe, N. Al Turkestani, K. Najarian et al., “Automatic multi-anatomical skull structure segmentation of cone-beam computed tomography scans using 3d unetr,” PLoS One , vol. 17, no. 10, p. e...

    2022

  19. [27]

    Dermatologist-level classification of skin cancer with deep neural networks,

    A. Esteva, B. Kuprel, R. A. Novoa, J. Ko, S. M. Swetter, H. M. Blau, and S. Thrun, “Dermatologist-level classification of skin cancer with deep neural networks,” nature, vol. 542, no. 7639, pp. 115–118, 2017

    2017

  20. [28]

    Artificial intelligence–based breast cancer nodal metastasis detection: insights into the black box for pathologists,

    Y . Liu, T. Kohlberger, M. Norouzi, G. E. Dahl, J. L. Smith, A. Mo- htashamian, N. Olson, L. H. Peng, J. D. Hipp, and M. C. Stumpe, “Artificial intelligence–based breast cancer nodal metastasis detection: insights into the black box for pathologists,” Archives of pathology & l...

    2019

  21. [29]

    Deeplensnet: deep learning automated diagnosis and quantitative classification of cataract type and severity,

    T. D. Keenan, Q. Chen, E. Agr ´on, Y .-C. Tham, J. H. L. Goh, X. Lei, Y . P. Ng, Y . Liu, X. Xu, C.-Y . Chenget al., “Deeplensnet: deep learning automated diagnosis and quantitative classification of cataract type and severity,” Ophthalmology, vol. 129, no. 5, pp. 571–584, 2022

    2022

  22. [30]

    A survey on image data augmen- tation for deep learning,

    C. Shorten and T. M. Khoshgoftaar, “A survey on image data augmen- tation for deep learning,” Journal of big data , vol. 6, no. 1, pp. 1–48, 2019

    2019

  23. [31]

    Cutmix: Reg- ularization strategy to train strong classifiers with localizable features,

    S. Yun, D. Han, S. J. Oh, S. Chun, J. Choe, and Y . Yoo, “Cutmix: Reg- ularization strategy to train strong classifiers with localizable features,” in Proceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 6023–6032. LIU et al.: GLIOMA MUL TIMODAL M...

    2019

  24. [32]

    U-net: deep learning for cell counting, detection, and morphometry,

    T. Falk, D. Mai, R. Bensch, ¨O. C ¸ ic ¸ek, A. Abdulkadir, Y . Marrakchi, A. B¨ohm, J. Deubner, Z. J¨ackel, K. Seiwald et al., “U-net: deep learning for cell counting, detection, and morphometry,” Nature methods, vol. 16, no. 1, pp. 67–70, 2019

    2019

  25. [33]

    A survey on vision transformer,

    K. Han, Y . Wang, H. Chen, X. Chen, J. Guo, Z. Liu, Y . Tang, A. Xiao, C. Xu, Y . Xuet al., “A survey on vision transformer,” IEEE transactions on pattern analysis and machine intelligence, vol. 45, no. 1, pp. 87–110, 2022

    2022

  26. [34]

    What uncertainties do we need in bayesian deep learning for computer vision?

    A. Kendall and Y . Gal, “What uncertainties do we need in bayesian deep learning for computer vision?” Advances in neural information processing systems, vol. 30, 2017

    2017

  27. [35]

    Missing mri pulse sequence synthesis using multi-modal generative adversarial network,

    A. Sharma and G. Hamarneh, “Missing mri pulse sequence synthesis using multi-modal generative adversarial network,” IEEE transactions on medical imaging , vol. 39, no. 4, pp. 1170–1183, 2019

    2019

  28. [36]

    Self-distillation: Towards efficient and compact neural networks,

    L. Zhang, C. Bao, and K. Ma, “Self-distillation: Towards efficient and compact neural networks,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 8, pp. 4388–4403, 2021

    2021

  29. [37]

    Learning from architectural redundancy: Enhanced deep supervision in deep multipath encoder– decoder networks,

    Y . Luo, J. L ¨u, X. Jiang, and B. Zhang, “Learning from architectural redundancy: Enhanced deep supervision in deep multipath encoder– decoder networks,” IEEE Transactions on Neural Networks and Learn- ing Systems, vol. 33, no. 9, pp. 4271–4284, 2021

    2021

  30. [38]

    A calibration hierarchy for risk models was defined: from utopia to empirical data,

    B. Van Calster, D. Nieboer, Y . Vergouwe, B. De Cock, M. J. Pencina, and E. W. Steyerberg, “A calibration hierarchy for risk models was defined: from utopia to empirical data,” Journal of clinical epidemiology, vol. 74, pp. 167–176, 2016

    2016

  31. [39]

    Remixmatch: Semi-supervised learning with distribution alignment and augmentation anchoring,

    D. Berthelot, N. Carlini, E. D. Cubuk, A. Kurakin, K. Sohn, H. Zhang, and C. Raffel, “Remixmatch: Semi-supervised learning with distribution alignment and augmentation anchoring,” arXiv preprint arXiv:1911.09785, 2019

    1911 arXiv

  32. [40]

    Unimatch v2: Pushing the limit of semi-supervised semantic segmentation,

    L. Yang, Z. Zhao, and H. Zhao, “Unimatch v2: Pushing the limit of semi-supervised semantic segmentation,” IEEE Transactions on Pattern Analysis and Machine Intelligence , 2025

    2025

  33. [41]

    Fixmatch: Simplifying semi- supervised learning with consistency and confidence,

    K. Sohn, D. Berthelot, N. Carlini, Z. Zhang, H. Zhang, C. A. Raffel, E. D. Cubuk, A. Kurakin, and C.-L. Li, “Fixmatch: Simplifying semi- supervised learning with consistency and confidence,” Advances in neural information processing systems , vol. 33, pp. 596–608, 2020

    2020

  34. [42]

    Semi-supervised semantic segmentation with cross pseudo supervision,

    X. Chen, Y . Yuan, G. Zeng, and J. Wang, “Semi-supervised semantic segmentation with cross pseudo supervision,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 2613–2622. 20 GENERIC COLORIZED JOURNAL, VOL. XX, NO. XX, 2025 VI. A ...

    2021

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.