Pith. sign in

REVIEW 5 major objections 6 minor 47 references

DMAF-Net: An Effective Modality Rebalancing Framework for Incomplete Multi-Modal Medical Image Segmentation

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read DMAF-Net rebalances unevenly missing MRI modalities and beats prior methods on two segmentation benchmarks.

desk verdict Solid, incremental engineering with modest real gains, but the DTM gradient-scaling mechanism is largely inert under AdamW and needs a clean ablation. read the letter →

arxiv 2506.11691 v1 pith:4L6EJVB6 submitted 2025-06-13 cs.CV

Libin Lan , Hongxing Li , Zunhui Xia , Yudong Zhang This is my paper · ORCID
classification cs.CV
keywords incompletemulti-modalsegmentationmodalityimbalancedynamicmodality-awarefusionrelationandprototypedistillationtrainingmonitoringbraintumorcardiacMRImissing
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 addresses incomplete multi-modal medical image segmentation in the clinically realistic setting where modalities are missing at unequal rates, a condition the authors call Imperfect Data Training. It proposes DMAF-Net, a framework with three coordinated mechanisms: a dynamic fusion module that masks missing modalities and weights available ones through transformer attention; a dual-level distillation scheme aligning global statistics, local attention, and class-specific prototypes between uni-modal and fused features; and a Dynamic Training Monitoring strategy that tracks real-time distillation gaps to reweight losses and scale gradients per modality. The paper claims DMAF-Net consistently outperforms existing modality-rebalancing methods, including PASSION, PMR, and ModDrop, on the BraTS2020 and MyoPS2020 benchmarks across every tested missing-rate configuration. If true, this provides a practical route to robust segmentation when clinical MRI data is incompletely acquired.

What carries the argument

The load-bearing object is the distillation-gap signal computed in the Dynamic Training Monitoring strategy: for each modality, the gap combines an EMA-smoothed relation distillation loss (covariance and masked-attention alignment) with a prototype distillation loss, producing an inverse weight that rescales the uni-modal segmentation loss and a gradient scaling factor for that modality's encoder. The Dynamic Modality-Aware Fusion module carries the input side of the argument by tokenizing each modality's features, applying a learned attention mask that zeroes missing modalities, and generating spatial weight maps for a weighted fusion. Relation distillation enforces global covariance consistency and local masked-attention alignment between fused and uni-modal features, while prototype distillation aligns class-specific prototypes to preserve semantic consistency. The three mechanisms work together to balance convergence speeds across modalities with heterogeneous missing rates.

What would settle it

Train DMAF-Net on BraTS2020 under missing rates (0.2, 0.4, 0.6, 0.8) with the DTM module disabled and all modality weights fixed to equal values; if average Dice does not drop markedly (for instance, by more than one point), the gap-based reweighting is not the source of the gain. A complementary test is to invert the mapping so the highest-gap modality receives the lowest weight; if segmentation quality is unaffected, the gap signal carries no causal information.

Watch

Extended reading notes

Core claim

The central claim is that explicitly modeling both imbalanced modality missing rates and heterogeneous modality contributions yields consistent segmentation improvements over methods that assume balanced, randomly masked training data. DMAF-Net achieves this through its three components acting in concert: attention-based dynamic fusion suppresses interference from absent modalities while weighting available ones; relation and prototype distillation transfers global statistical structure and category-level semantics from fused to uni-modal features; and dynamic training monitoring uses the model's own distillation gaps as a real-time proxy for each modality's learning deficit, adaptively reweighting losses and scaling gradients to slow dominant modalities and accelerate under-trained ones. The authors report that the full framework outperforms state-of-the-art modality rebalancing methods on BraTS2020 and MyoPS2020 under multiple imbalanced missing-rate settings, with ablation studies attributing the best results to joint integration of all three components.

Load-bearing premise

The Dynamic Training Monitoring strategy assumes the distillation gap measured from the model's own uni-modal and fused features is a reliable, real-time indicator of each modality's true learning deficit; if random masking noise or dominance of the strongest modality corrupts that gap, the adaptive reweighting and gradient scaling would misallocate capacity and the reported improvements could disappear.

Editorial extensions

If this is right

  • Under imbalanced missing rates, DMAF-Net reports higher Dice and lower Hausdorff distances than ModDrop, PMR, and PASSION on both BraTS2020 and MyoPS2020 in every tested configuration.
  • The joint ablation shows each component contributes, and the full model outperforms any single component or pair, indicating the mechanisms are complementary rather than redundant.
  • Because DMAF-Net keeps whole-tumor segmentation under imperfect data training close to its perfect-data-training level, the framework reduces the performance penalty imposed by uneven modality absence.
  • The architecture transfers across 3D (BraTS2020) and 2D (MyoPS2020) segmentation, suggesting the rebalancing mechanisms are not tied to a specific dimensionality or organ system.

Reading between the lines

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

  • The real-time distillation-gap idea could generalize beyond segmentation: any multi-modal task with heterogeneous modality quality or acquisition cost could use a similar gap signal to pace training per modality.
  • A practical extension would be to expose the distillation gap as an online confidence estimate for each modality, flagging scans whose missing sequences most degrade fusion quality for downstream quality control.
  • The gradient-conflict heuristic (0.7 suppression when consecutive gradient directions are anti-aligned) is chosen by grid search, so a principled derivation of the conflict threshold could make the monitoring strategy less dataset-dependent.
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

5 major / 6 minor

Summary. This paper proposes DMAF-Net, a framework for incomplete multi-modal medical image segmentation under imbalanced modality missing rates. It combines a Dynamic Modality-Aware Fusion (DMAF) module based on masked multi-head attention with spatial attention weighting, a dual-level relation and prototype distillation scheme, and a Dynamic Training Monitoring (DTM) strategy that adaptively reweights modality-specific losses and scales encoder gradients based on EMA-smoothed distillation gaps. The method is evaluated on BraTS2020 (3D) and MyoPS2020 (2D) under various missing-rate settings, reporting improved Dice and Hausdorff metrics over the mmFormer baseline and several rebalancing methods (ModDrop, PMR, PASSION), with ablations in Table 3.

Significance. If the reported results are reliable, DMAF-Net would be a competitive approach for incomplete multimodal segmentation under heterogeneous missing rates, and the design ideas (attention-masked fusion, covariance/attention distillation, and gap-based training monitoring) are interesting. The paper has clear strengths: code is released, the evaluation covers multiple missing-rate patterns on two datasets, and the ablation isolates the three components. However, the claimed improvements are modest and lack statistical substantiation, and the core DTM mechanism has a mathematical ambiguity and a potential optimizer interaction that weakens the mechanistic story. The contribution is incremental rather than a leap, but if strengthened it could be of interest to the field.

major comments (5)
  1. [Section 4.2 and Table 1] No error bars or statistical significance tests are reported. Across the IDT settings in Table 1, the DSC differences between DMAF-Net and PASSION are typically below 1.5 percentage points (e.g., BraTS2020 Avg 69.83 vs 68.76; 69.98 vs 69.05; etc.). Without multiple seeds or significance testing, the central 'outperforms' claim is not established. The authors should report mean±std over at least 3 seeds and, where possible, paired significance tests.
  2. [Section 4.3 and Table 1] The claim of 'consistent performance improvements across all missing rate scenarios' is contradicted by the (l,s,m) row for BraTS2020, where DMAF-Net's WT DSC (82.43) is lower than PASSION's (83.12). The abstract's unqualified statement that DMAF-Net 'outperforms existing methods' should be qualified to reflect this exception.
  3. [Section 3.4, Eq. (18) with Section 4.2] The DTM gradient-scaling mechanism is described as decelerating dominant modalities and accelerating underrepresented ones, but the optimizer is AdamW, which normalizes gradients by their second moment. For a slowly varying scalar gamma(t)(m), the effective parameter update under Adam is nearly invariant to the scale of gamma. Therefore the claimed balancing effect of Eq. (18) is largely inert; only the loss reweighting in Eq. (17) can alter the objective. The paper should either replace AdamW with an optimizer without per-parameter normalization (or otherwise derive the update), or provide an ablation that disentangles loss reweighting from gradient scaling to show the gradient scaling contributes.
  4. [Eq. (11) of Section 3.4] The prototype distillation gap g_p(m) is defined with a single, unspecified class c, whereas the prototype loss in Eq. (10) sums over all classes. This makes the DTM weight computation (Eqs. 13-16) ambiguous for multi-class segmentation. It should be clarified whether g_p is the class-average, class-sum, or maximum over classes.
  5. [Section 2.2 and Tables 1-2] The related-work discussion mentions RedCore [19] and DynCIM [40] as directly addressing imbalanced missing rates, but neither method appears in the comparisons. Since both were available before submission, the claim of outperforming state-of-the-art rebalancing methods is not fully supported. The authors should add these comparisons or justify their exclusion.
minor comments (6)
  1. [Table 2, TC/IDT row] The entry '9.17' for DMAF-Net is an obvious typo and should be roughly 69; this error undermines confidence in the other reported numbers.
  2. [Abstract] The sentence 'Extensive experiments on BraTS2020 and MyoPS2020 demonstrate that DMAF-Net outperforms existing methods for incomplete multi-modal medical image segmentation.' appears twice; remove the duplicate.
  3. [Eq. (8)] The symbol L^m_conv is used where the surrounding text and Eq. (5) refer to the covariance loss L^m_cov; the notation should be made consistent.
  4. [Eq. (10) and Eq. (11)] There are mismatched parentheses: '(1 - cos(p^f_c, p^m_c)/tau^m)' presumably means '(1 - cos(p^f_c, p^m_c)) / tau^m'. Please clarify the intended temperature scaling.
  5. [Table 3 header] The header spells 'DISTINATION' instead of 'DISTILLATION'.
  6. [References] The citation [24] appears to be a modality-gap classification paper, while the text uses it for HEALNet's hybrid early fusion; please verify the reference and possibly correct the citation.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's central claim is an empirical benchmark result against external baselines on public datasets, not a derivation forced by fitted parameters or self-citations.

full rationale

The central claim of DMAF-Net is empirical: Tables 1-3 report DSC/HD on BraTS2020 and MyoPS2020 against external baselines (mmFormer, RFNet, M2FTrans) and rebalancing methods (ModDrop, PMR, PASSION) under identical missing-rate protocols. No section derives the reported improvements from the method's own equations; the comparisons are self-contained and falsifiable. The DTM strategy (Section 3.4) defines its 'distillation gaps' in Eq. (11) as the same per-modality relation and prototype distillation losses from Eqs. (8) and (10), and then reweights losses and scales gradients as functions of those gaps. This is a self-referential training feedback loop, but it is a heuristic update rule, not a derivation of the claimed gain; the paper's conclusion rests on external experiments and ablations, not on a logical consequence of the gap definitions. Grid-searched hyperparameters (alpha1, lambda1-4, the 0.7 conflict factor in Section 4.2 and Eq. (18)) are tuned on the same datasets, which is a generalization concern, but no reported metric is a fitted parameter renamed as a prediction. Even if Reference [19] includes a current author, it is cited only as related work and is not load-bearing; no uniqueness theorem or prior author result is invoked to force the architecture. The AdamW gradient-scaling concern raised in review is a correctness or effectiveness issue, not circularity. The paper's own limitations section acknowledges computational cost and prototype vulnerability to label noise, further showing that the claims are treated as empirical rather than derivational.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

The central claim rests on a set of tuned hyperparameters and standard architectural assumptions. No new physical entities or independent theoretical objects are introduced. The DTM gap mechanism is self-referential but not circular in a logical sense.

free parameters (7)
  • alpha1 = 0.6
    Balancing weight between covariance and attention losses in relation distillation (Sec 4.2). Selected by grid search.
  • lambda1, lambda2, lambda3, lambda4 = 2.0, 1.0, 0.5, 0.5
    Weights for the four loss terms in Eq. (19). Selected by grid search (Sec 4.2).
  • Gradient conflict scaling factor = 0.7
    Factor applied when consecutive gradient cosine similarity is below -0.5 (Eq. 18). Chosen by grid search over {0.5,0.6,0.7,0.8}.
  • Gradient scaling clip range = [0.1, 10.0]
    Hard bounds on gamma in Eq. (18). Not justified except as a design choice.
  • Cosine similarity threshold = -0.5
    Threshold in Eq. (18) for detecting conflicting gradient directions. Not further justified.
  • EMA decay base = 0.9
    Base decay in Eq. (13) adjusted by sigmoid of gap ratio. Chosen without exposed sensitivity analysis.
  • Prototype temperature tau_m = not specified
    Temperature in prototype loss (Eq. 10) controls per-modality optimization speed, but no value or update rule is provided.
assumptions (4)
  • standard math Transformer attention and covariance matrix operations behave as defined in standard deep learning.
    The method relies on standard MMHA and covariance computations without formal proofs.
  • domain assumption Each sample contains at least one modality and all available modalities share identical segmentation labels.
    Stated in Section 3.1, Problem Definition. This underpins the multi-encoder architecture and the fusion objective.
  • domain assumption The modality presence matrix I is known at both training and test time and accurately encodes missingness.
    The DMAF mask and DTM gaps depend on accurate knowledge of which modalities are present, which is a reasonable but unstated assumption for deployment.
  • ad hoc to paper The gradient conflict threshold (-0.5) and the 0.7 scaling factor, tuned on the same benchmarks, will transfer to other datasets.
    These constants are selected by grid search on BraTS2020 and MyoPS2020 (Eq. 18 description), and the paper provides no evidence of robustness across other datasets.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DMAF-Net: An Effective Modality Rebalancing Framework for Incomplete Multi-Modal Medical Image Segmentation." pith.science (2026). https://pith.science/paper/4L6EJVB6

@misc{pith2026250611691,
  author       = {Pith},
  title        = {Pith review of: DMAF-Net: An Effective Modality Rebalancing Framework for Incomplete Multi-Modal Medical Image Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4L6EJVB6}},
  note         = {Machine review of arXiv:2506.11691}
}
read the original abstract

Incomplete multi-modal medical image segmentation faces critical challenges from modality imbalance, including imbalanced modality missing rates and heterogeneous modality contributions. Due to their reliance on idealized assumptions of complete modality availability, existing methods fail to dynamically balance contributions and neglect the structural relationships between modalities, resulting in suboptimal performance in real-world clinical scenarios. To address these limitations, we propose a novel model, named Dynamic Modality-Aware Fusion Network (DMAF-Net). The DMAF-Net adopts three key ideas. First, it introduces a Dynamic Modality-Aware Fusion (DMAF) module to suppress missing-modality interference by combining transformer attention with adaptive masking and weight modality contributions dynamically through attention maps. Second, it designs a synergistic Relation Distillation and Prototype Distillation framework to enforce global-local feature alignment via covariance consistency and masked graph attention, while ensuring semantic consistency through cross-modal class-specific prototype alignment. Third, it presents a Dynamic Training Monitoring (DTM) strategy to stabilize optimization under imbalanced missing rates by tracking distillation gaps in real-time, and to balance convergence speeds across modalities by adaptively reweighting losses and scaling gradients. Extensive experiments on BraTS2020 and MyoPS2020 demonstrate that DMAF-Net outperforms existing methods for incomplete multi-modal medical image segmentation. Extensive experiments on BraTS2020 and MyoPS2020 demonstrate that DMAF-Net outperforms existing methods for incomplete multi-modal medical image segmentation. Our code is available at https://github.com/violet-42/DMAF-Net.

Figures

Figures reproduced from arXiv: 2506.11691 by the authors.

Figure 1
Figure 1. Our proposed Dynamic Modality-Aware Fusion (DMAF) module. Here, 𝑅𝑓 denotes the feature dimensions (𝐷 × 𝐻 × 𝑊 for 3D volumes or 𝐻 × 𝑊 for 2D slices), while 𝑇 represents the tokenized form (with 𝐷𝐻𝑊 or 𝐻𝑊 tokens) suitable for attention computation in the transformer architecture. 𝑒 𝑚 𝑛,𝑙 indicates the output feature of the 𝑙-th layer in encoder 𝑚. causing them to underperform uni-modal approaches in spe￾cific diagnost… view at source ↗
Figure 2
Figure 2. Architecture of our proposed DMAF-Net. The framework processes multi-modal inputs through modality-specific encoders to extract individual feature representations. These features are processed via two distinct decoding pathways: (1) A shared uni-modal decoder that independently handles each modality’s features, and (2) a multi-modal fusion pathway where encoder outputs are dynamically integrated at each level throug… view at source ↗
Figure 3
Figure 3. Segmentation visualization with MR=0.2/0.4/0.6/0.8 (T1/T1c/FLAIR/T2) on BraTS2020. Panels (a-d): T2, T1c+T2, T1c+FLAIR+T2, and all four modalities. information disparity among modalities with closer seman￾tic contributions, eliminating dominant modality effects. Consequently, the impact of imbalanced modality missing rates on segmentation performance is relatively diminished. As a result, all modality rebalancing me… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Segmentation visualization with MR=0.3/0.5/0.7 (bSSFP/LGE/T2) on MyoPS2020. Panels (a-d): T2, LGE, bSSFP+T2, and bSSFP+LGE+T2. heterogeneity — exhibited significant performance degra￾dation under IDT (Imperfect Data Training) conditions, de￾spite their strong performan…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

47 extracted references · 43 canonical work pages

  1. [19]

    Z. Li, Y. Zhang, H. Li, Y. Chai, Y. Yang, Deformation-aware and reconstruction-driven multimodal representation learning for brain tumor segmentation with missing modalities, BSPC 91 (2024) 106012

    work page 2024

  2. [40]

    C. Qian, K. Han, J. Wang, Z. Yuan, R. Qian, C. Lyu, J. Chen, Z. Liu, Dyncim: Dynamic curriculum for imbalanced multimodal learning, arXiv preprint arXiv:2503.06456 (2025)

    arXiv 2025

  3. [1]

    Bakas, H

    S. Bakas, H. Akbari, A. Sotiras, M. Bilello, M. Rozycki, J. S. Kirby, J. B. Freymann, K. Farahani, C. Davatzikos, Advancing the cancer genome atlas glioma mri collections with expert segmentation labels and radiomic features, Sci. Data. 4 (1) (2017) 1–13

    work page 2017

  4. [2]

    P.Biondetti,M.G.Vangel,R.M.Lahoud,F.S.Furtado,B.R.Rosen, D.Groshar,L.G.Canamaque,L.Umutlu,E.W.Zhang,U.Mahmood, et al., Pet/mri assessment of lung nodules in primary abdominal ma- lignancies: sensitivity and outcome analysis, Eur. J. Nucl. Med. Mol. I. 48 (2021) 1976–1986

    work page 2021

  5. [3]

    J. Chen, J. Cui, Y. Yu, X. Wang, Magnetic resonance imaging vol- umesegmentationforbreastcancersurgery,Biomed.SignalProcess. Control. 110 (2025) 108131

    work page 2025

  6. [4]

    Y. Ding, X. Yu, Y. Yang, Rfnet: Region-aware fusion network for in- complete multi-modal brain tumor segmentation, in: Proc. IEEE Int. Conf. Comput. Vision., 2021, pp. 3975–3984

    work page 2021

  7. [5]

    G. Ma, X. Qiu, X. Tan, Dmfusion: A dual-branch multi-scale feature fusionnetworkformedicalmulti-modalimagefusion,Biomed.Signal Process. Control. 105 (2025) 107572

    work page 2025

  8. [6]

    Krupa, M

    K. Krupa, M. Bekiesińska-Figatowska, Artifacts in magnetic reso- nance imaging, POL. J. RADIOL 80 (2015) 93

    work page 2015

Show all 47 references
  1. [7]

    M.J.Graves,D.G.Mitchell,Bodymriartifactsinclinicalpractice:a physicist’s and radiologist’s perspective, J. MAGN. RESON. IMAG- ING. 38 (2) (2013) 269–287

    2013

  2. [8]

    Havaei, N

    M. Havaei, N. Guizard, N. Chapados, Y. Bengio, Hemis: Hetero-modal image segmentation, in: Proc. Int. Conf. Med. Image Comput. Comput.-Assisted Intervention., Springer, 2016, pp. 469–477

    2016

  3. [9]

    Comput.-Assisted Intervention., Springer, 2023, pp

    Z.Wang,Y.Hong,A2fseg:Adaptivemulti-modalfusionnetworkfor medicalimagesegmentation,in:Proc.Int.Conf.Med.ImageComput. Comput.-Assisted Intervention., Springer, 2023, pp. 673–681

    2023

  4. [10]

    J. Shi, C. Shang, Z. Sun, L. Yu, X. Yang, Z. Yan, Passion: Towards effective incomplete multi-modal medical image segmentation with imbalancedmissingrates,in:Proc.ACMInt.Conf.Multimed.,2024, pp. 456–465. L. Lan et al.:Preprint submitted to Elsevier Page 11 of 12

    2024

  5. [11]

    Goodfellow, J

    I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, Y. Bengio, Generative adversarial networks, COMMUN. ACM. 63 (11) (2020) 139–144

    2020

  6. [12]

    X. Meng, K. Sun, J. Xu, X. He, D. Shen, Multi-modal modality-masked diffusion network for brain mri synthesis with random modality missing, IEEE Trans. Med. Imag. (2024)

    2024

  7. [13]

    Müller-Franzes, J

    G. Müller-Franzes, J. M. Niehues, F. Khader, S. T. Arasteh, C. Haar- burger,C.Kuhl,T.Wang,T.Han,T.Nolte,S.Nebelung,etal.,Amul- timodalcomparisonoflatentdenoisingdiffusionprobabilisticmodels andgenerativeadversarialnetworksformedicalimagesynthesis,Sci. Rep. 13 (1) (2023) 12098

    2023

  8. [14]

    Zhang, C

    Y. Zhang, C. Peng, Q. Wang, D. Song, K. Li, S. K. Zhou, Unified multi-modal image synthesis for missing modality imputation, IEEE Trans. Med. Imag. (2024)

    2024

  9. [15]

    Zhang, G

    Y. Zhang, G. Liu, R. Bao, L. Zhan, P. Thompson, H. Huang, Brain image synthesis using incomplete multimodal data, in: Proc. IEEE Int. Symp. Biomed. Imaging, IEEE, 2024, pp. 1–5

    2024

  10. [16]

    IEEEInt.Conf.Acoust.SpeechSignalProcess.,IEEE,2023,pp.1–5

    S.Wang,Z.Yan,D.Zhang,H.Wei,Z.Li,R.Li,Prototypeknowledge distillationformedicalsegmentationwithmissingmodality,in:Proc. IEEEInt.Conf.Acoust.SpeechSignalProcess.,IEEE,2023,pp.1–5

    2023

  11. [17]

    H. Wang, C. Ma, J. Zhang, Y. Zhang, J. Avery, L. Hull, G. Carneiro, Learnable cross-modal knowledge distillation for multi-modal learn- ing with missing modality, in: Proc. Int. Conf. Med. Image Comput. Comput.-Assisted Intervention., Springer, 2023, pp. 216–226

    2023

  12. [18]

    Zhang, N

    Y. Zhang, N. He, J. Yang, Y. Li, D. Wei, Y. Huang, Y. Zhang, Z. He, Y. Zheng, mmformer: Multimodal medical transformer for incomplete multimodal learning of brain tumor segmentation, in: Proc.Int.Conf.Med.ImageComput.Comput.-AssistedIntervention., Springer, 2022, pp. 107–117

    2022

  13. [20]

    15878–15887

    H.Wang,Y.Chen,C.Ma,J.Avery,L.Hull,G.Carneiro,Multi-modal learningwithmissingmodalityviashared-specificfeaturemodelling, in:Proc.IEEEInt.Conf.Comput.VisionPatternRecognit.,2023,pp. 15878–15887

    2023

  14. [21]

    Y.Wei,D.Hu,H.Du,J.-R.Wen,On-the-flymodulationforbalanced multimodallearning,IEEETrans.PatternAnal.Mach.Intell.(2024)

    2024

  15. [22]

    X. Peng, Y. Wei, A. Deng, D. Wang, D. Hu, Balanced multimodal learningviaon-the-flygradientmodulation,in:Proc.IEEEInt.Conf. Comput. Vision Pattern Recognit., 2022, pp. 8238–8247

    2022

  16. [23]

    Y. Fan, W. Xu, H. Wang, J. Wang, S. Guo, Pmr: Prototypical modal rebalanceformultimodallearning,in:Proc.IEEEInt.Conf.Comput. Vision Pattern Recognit., 2023, pp. 20029–20038

    2023

  17. [24]

    Y. Yang, F. Wan, Q.-Y. Jiang, Y. Xu, Facilitating multimodal classifi- cation via dynamically learning modality gap, Proc. Adv. Neural Inf. Proces. Syst. 37 (2024) 62108–62122

    2024

  18. [25]

    Zhang, J

    X. Zhang, J. Yoon, M. Bansal, H. Yao, Multimodal representation learningbyalternatingunimodaladaptation,in:Proc.IEEEInt.Conf. Comput. Vision Pattern Recognit., 2024, pp. 27456–27466

    2024

  19. [26]

    IEEE Int

    Y.Wei,R.Feng,Z.Wang,D.Hu,Enhancingmultimodalcooperation via sample-level modality valuation, in: Proc. IEEE Int. Conf. Com- put. Vision Pattern Recognit., 2024, pp. 27338–27347

    2024

  20. [27]

    Zhang, Z

    Y. Zhang, Z. Chen, L. Liang, H. Chen, W. Zhang, Unleashing the power of imbalanced modality information for multi-modal knowl- edge graph completion, in: Proc. Jt. Int. Conf. Comput. Linguist., 2024, pp. 17120–17130

    2024

  21. [28]

    AAAI Conf

    J.Sun,X.Zhang,S.Han,Y.-P.Ruan,T.Li,Redcore:Relativeadvan- tageawarecross-modalrepresentationlearningformissingmodalities with imbalanced missing rates, in: Proc. AAAI Conf. Artif. Intell., Vol. 38, 2024, pp. 15173–15182

    2024

  22. [29]

    J. Ho, A. Jain, P. Abbeel, Denoising diffusion probabilistic models, Proc. Adv. Neural Inf. Proces. Syst. 33 (2020) 6840–6851

    2020

  23. [30]

    M. Ma, J. Ren, L. Zhao, S. Tulyakov, C. Wu, X. Peng, Smil: Multi- modallearningwithseverelymissingmodality,in:Proc.AAAIConf. Artif. Intell., Vol. 35, 2021, pp. 2302–2310

    2021

  24. [31]

    Xie, Motion artifact removal in coronary ct angiography based on generative adversarial networks, EUR

    L.Zhang,B.Jiang,Q.Chen,L.Wang,K.Zhao,Y.Zhang,R.Vliegen- thart, X. Xie, Motion artifact removal in coronary ct angiography based on generative adversarial networks, EUR. RADIOL. 33 (1) (2023) 43–53

    2023

  25. [32]

    Ahmad, H

    W. Ahmad, H. Ali, Z. Shah, S. Azmat, A new generative adversarial network for medical images super resolution, Sci. Rep. 12 (1) (2022) 9533

    2022

  26. [33]

    A. Wu, J. Yu, Y. Wang, C. Deng, Prototype-decomposed knowledge distillation for learning generalized federated representation, IEEE Trans. Multimedia (2024)

    2024

  27. [34]

    S. Wei, C. Luo, Y. Luo, Mmanet: Margin-aware distillation and modality-aware regularization for incomplete multimodal learning, in: Proc. IEEE Int. Conf. Comput. Vision Pattern Recognit., 2023, pp. 20039–20049

    2023

  28. [35]

    F. Xiao, Y. J. Lee, K. Grauman, J. Malik, C. Feichtenhofer, Au- diovisual slowfast networks for video recognition, arXiv preprint arXiv:2001.08740 (2020)

    2020 arXiv

  29. [36]

    Q. Dou, Q. Liu, P. A. Heng, B. Glocker, Unpaired multi-modal seg- mentation via knowledge distillation, IEEE Trans. Med. Imag. 39 (7) (2020) 2415–2425

    2020

  30. [37]

    Salguero, P

    J. Salguero, P. Prasanna, G. Corredor, A. Cruz-Roa, D. Becerra, E. Romero, Data distillation in computational pathology by choos- ing few representants of the original variance: A use case in ovarian cancer, EXPERT. SYST. APPL. 245 (2024) 123028

    2024

  31. [38]

    K. Wang, F. Zheng, D. Guan, J. Liu, J. Qin, Distilling heterogeneous knowledgewithalignedbiologicalentitiesforhistologicalimageclas- sification, Pattern Recognit. 160 (2025) 111173

    2025

  32. [39]

    J. Shi, L. Yu, Q. Cheng, X. Yang, K.-T. Cheng, Z. Yan, Mftrans: Modality-masked fusion transformer for incomplete multi-modality brain tumor segmentation, IEEE J. Biomed. Health Inform. 28 (1) (2023) 379–390

    2023

  33. [41]

    H. Li, X. Li, P. Hu, Y. Lei, C. Li, Y. Zhou, Boosting multi-modal modelperformancewithadaptivegradientmodulation,in:Proc.IEEE Int. Conf. Comput. Vision., 2023, pp. 22214–22224

    2023

  34. [42]

    Y.Wei,S.Li,R.Feng,D.Hu,Diagnosingandre-learningforbalanced multimodal learning, in: Proc. Eur. Conf. Comput. Vis., Springer, 2024, pp. 71–86

    2024

  35. [43]

    Morales-Brotons, T

    D. Morales-Brotons, T. Vogels, H. Hendrikx, Exponential moving average of weights in deep learning: Dynamics and benefits, arXiv preprint arXiv:2411.18704 (2024)

    2024 arXiv

  36. [44]

    B. H. Menze, A. Jakab, S. Bauer, J. Kalpathy-Cramer, K. Farahani, J. Kirby, Y. Burren, N. Porz, J. Slotboom, R. Wiest, et al., The mul- timodal brain tumor image segmentation benchmark (brats), IEEE Trans. Med. Imag. 34 (10) (2014) 1993–2024

    2014

  37. [45]

    J. Qiu, L. Li, S. Wang, K. Zhang, Y. Chen, S. Yang, X. Zhuang, Myops-net: Myocardial pathology segmentation with flexible com- bination of multi-sequence cmr images, MED. IMAGE. ANAL. 84 (2023) 102694

    2023

  38. [46]

    Zhuang, Multivariate mixture model for myocardial segmentation combiningmulti-sourceimages,IEEETrans.PatternAnal.Mach.In- tell

    X. Zhuang, Multivariate mixture model for myocardial segmentation combiningmulti-sourceimages,IEEETrans.PatternAnal.Mach.In- tell. 41 (12) (2018) 2933–2946

    2018

  39. [47]

    Loshchilov, F

    I. Loshchilov, F. Hutter, Decoupled weight decay regularization, arXiv preprint arXiv:1711.05101 (2017). L. Lan et al.:Preprint submitted to Elsevier Page 12 of 12

    2017 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.