Pith. sign in

REVIEW 4 major objections 4 minor 39 references

AdaMM, a knowledge-distillation framework with three synergistic modules, claims consistently superior brain-tumor segmentation under missing MRI modalities, beating 15 existing methods across BraTS 2018 and 2024.

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

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

T0 review · deepseek-v4-flash

2026-08-04 16:12 UTC pith:KYY2TC4I

load-bearing objection Useful method and benchmark, but the 'consistent superiority' claim rests on single-run, unequal-effort comparisons; needs multi-seed evidence and fixes to citation and unit errors. the 4 major comments →

arxiv 2509.15017 v2 pith:KYY2TC4I submitted 2025-09-18 cs.CV

No Modality Left Behind: Adapting to Missing Modalities via Knowledge Distillation for Brain Tumor Segmentation

Shenghao Zhu , Yifei Chen , Weihong Chen , Shuo Jiang , Guanyu Zhou , Yuanhan Wang , Feiwei Qin , Changmiao Wang
show 1 more author
Qiyuan Tian
This is my paper
classification cs.CV
keywords missing-modality MRIbrain tumor segmentationknowledge distillationgraph neural networkmodality adaptationlesion-presence priorBraTSmulti-modal learning
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper sets out to solve a clinical reality: multi-modal MRI for brain tumor segmentation frequently arrives incomplete. It proposes AdaMM, a teacher-student framework in which a full-modality teacher distills structural and textural knowledge into a student that must segment from whatever subset of scans is present. Three modules carry the work: a graph-guided refinement module models associations between generalizable and modality-specific features; a bi-bottleneck distillation module aligns teacher-student style and features; and a lesion-presence module uses predicted tumor-type probabilities to suppress false positives. Across 15 modality combinations on BraTS 2018 and 2024, AdaMM reports higher Dice, IoU, Sensitivity and lower HD95 than 15 comparison methods, with the largest gains in single-modality and weak-modality settings. The paper also positions knowledge distillation as the strongest of six methodological families for missing-modality segmentation.

Core claim

The central claim is that missing-modality brain-tumor segmentation is best handled not by imputing absent scans or features, but by knowledge distillation from a complete-modality teacher to a partial-modality student. The student uses an Adapter Bank with one small 3D residual module per modality combination, a graph network that ties combination-specific features to generalizable features, and a lesion-presence classifier that gates out unlikely tumor regions. On BraTS 2018 and 2024, the method reports consistent superiority, with average Dice gains of 23.55 and 23.74 percentage points over data-generation baselines in single-modality settings, and lower HD95 than feature-generation metho

What carries the argument

The framework's load-bearing objects are the three modules: GARM (Graph-guided Adaptive Refinement Module) uses learnable graph nodes and soft voxel-to-node assignments to relate modality-combination-specific adapter features to general features; BBDM (Bi-Bottleneck Distillation Module) computes global style matchings from encoder, decoder and bottleneck features and adds a discriminator to align teacher-student distributions; LGRM (Lesion-Presence-Guided Reliability Module) predicts a per-lesion presence probability and uses it to weight Dice loss and to zero out absent tumor classes at inference. Together they convert missing-modality robustness into a distillation problem with explicit st

Load-bearing premise

The comparison assumes the 15 baseline methods were trained and tuned with the same effort and protocol as AdaMM; the paper reports hyperparameters only for its own model and gives no variance across repeated runs.

What would settle it

Retrain all 15 baselines and AdaMM with identical hyperparameter search budgets and multiple seeds on the same BraTS splits; if the average Dice differences between AdaMM and the best baseline fall below roughly one standard deviation, or if any baseline matches AdaMM in single-modality settings, the claim of consistent superiority is falsified.

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

If this is right

  • If AdaMM's results hold, clinical deployments can segment tumors from whatever MRI sequences are available, rather than requiring a full four-sequence protocol.
  • The six-family comparison suggests researchers should prioritize knowledge distillation over GAN/VAE imputation for missing-modality segmentation, saving compute and avoiding synthetic-image artifacts.
  • The per-combination adapter bank plus graph alignment offers a recipe for adapting a single model to many input subsets without retraining per subset.
  • The lesion-presence prior directly targets false positives, which matters for treatment planning where spurious regions could mislead radiotherapy dosage.

Where Pith is reading between the lines

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

  • The adapter-bank design has an implicit combinatorial cost: one adapter per modality combination (15 here), which could become unwieldy as modality count grows; a more scalable parameter-sharing scheme would be a natural extension the paper does not explore.
  • The claimed advantage over imputation could be tested on other multi-modal medical tasks (e.g., pathology-radiology or multi-contrast CT) where missing inputs are common; the mechanism is task-agnostic in principle.
  • Since no repeated-seed variance is reported, a fair re-benchmark that tunes all baselines with equal budget and reports error bars would be the decisive check on 'consistent superiority'.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes AdaMM, a knowledge-distillation-based framework for brain tumor segmentation under missing MRI modalities. It combines a Graph-guided Adaptive Refinement Module (GARM), a Bi-Bottleneck Distillation Module (BBDM), and a Lesion-Presence-Guided Reliability Module (LGRM). The authors evaluate on BraTS 2018 and BraTS 2024 across 15 modality-missing configurations, comparing against 15 methods from six methodological categories. The central claim is that AdaMM consistently outperforms existing methods, especially in single- and weak-modality settings, and that knowledge distillation is generally superior for this task. The paper also includes ablations and qualitative visualizations.

Significance. If the reported superiority is reliable, AdaMM would be a practically useful contribution to missing-modality brain tumor segmentation, with the systematic six-category comparison providing a useful benchmark. The release of source code is a positive element. However, the significance is currently undermined by evaluation weaknesses: the key comparative claim rests on single-run point estimates without variance or significance testing, the baseline training protocol is not documented, and several tables contain inconsistencies. These issues need to be addressed before the claims can be accepted.

major comments (4)
  1. [Section 4.4, Tables 1–8] The central claim of consistent superiority over 15 baselines relies entirely on single-run point estimates in Tables 1–8. No standard deviations, confidence intervals, or statistical significance tests are reported. The BraTS 2018 test split is only about 57 cases, so Dice differences of a few points, typical of run-to-run and initialization variance, could alter rankings. Section 4.2 specifies hyperparameters (epochs, learning rate, input size) only for the proposed model and gives no training protocol for the baselines. Without matched tuning and repeated runs, the claimed consistent improvement cannot be distinguished from an artifact of asymmetric evaluation effort. Please add multi-seed experiments, report mean±std, and perform paired significance tests on the main comparisons.
  2. [Table 7, TC row for Hyper-GAE] In the BraTS 2018 IoU table, the Hyper-GAE TC row reports values such as 0.44, 0.47, 0.66, while all other entries in the same table are between roughly 20 and 80 on a percentage scale. The corresponding DSC values in Table 5 are around 50–80, making these IoU values implausible. This indicates a unit or reporting error and undermines confidence in the accuracy of the tabulated results. Please recheck and regenerate all tables, ensuring consistent units.
  3. [References [34] and [35]] The baseline named SMUNet [35] is described as a style-matching UNet for missing-modality brain tumor segmentation, but the cited reference is 'SMU-Net: Saliency-guided morphology-aware U-Net for breast lesion segmentation in ultrasound image' (IEEE TMI, 2021). That paper is not a style-matching UNet for brain tumors. If the authors implemented a different style-matching UNet, they must provide the correct citation and implementation details; otherwise the comparison is against a misidentified baseline. This also raises concerns about the fidelity of other baseline implementations.
  4. [Abstract and Section 4.1] The abstract states that experiments were conducted on 'Pretreat-MetsToBrain-Masks' in addition to BraTS 2018 and 2024. However, Section 4.1 describes only BraTS 2018 and BraTS 2024, and none of the tables include a third dataset. This inconsistency must be resolved: if the third dataset was used, report its results; if not, remove it from the abstract.
minor comments (4)
  1. [Section 4.5.1–4.5.3] The ablation text appears to mis-order the Dice decreases. For example, in Section 4.5.1 the text says 'ET, WT, and TC fell by 5.16%, 3.01%, and 2.82%', but Table 1 shows WT fell by 5.16, TC by 3.01, and ET by 2.82. Similar mismatches occur in Sections 4.5.2 and 4.5.3. Please align the reported order with the tables.
  2. [Equation (2)] The normalization factor θ/(4n²) is unexplained. Please define n and the origin of the factor 4, or simplify the notation.
  3. [Figures 3 and 4] The architecture diagrams are dense and low-resolution. Labels such as ℒM1&3 and ℒM1&2 are hard to read. A higher-resolution version with separated loss paths would improve readability.
  4. [Section 3.4 and Figure 2] Typo: 'to to improve' should be 'to improve'. The Figure 2 caption says 'Muti-task Learning'; should be 'Multi-task Learning'.

Circularity Check

0 steps flagged

No circularity: teacher–student distillation, graph refinement, and lesion-presence gating are all trained on data and evaluated against ground truth; self-citations are disclosed and not load-bearing.

full rationale

The derivation chain is empirical and self-contained: AdaMM's losses (Eq. 20) combine a teacher–student MSE distillation term, the BBDM style/adversarial alignment, and the LGRM classification/segmentation losses. The student is trained on incomplete inputs and evaluated with Dice/IoU/HD95/Sensitivity against manual ground-truth masks, not against the teacher's output or any fitted parameter. The LGRM auxiliary presence classifier is trained with labels obtained by max-pooling ground truth (Eq. 15-18), and at inference the gating uses the predicted presence probability, not the ground-truth label; this is a standard auxiliary-task design, not a self-definitional loop. The self-citations ([33], [36]) are previous works by the authors used as related work and as one of the 15 baselines; they are not invoked to justify the architecture, to prove uniqueness, or to define the evaluation metric. The main weaknesses are experimental rather than circular: baseline training hyperparameters are not reported, tables show single runs without variance or significance tests, and reference [35] is mismatched (cited as 'Style-Matching UNet' but points to a breast-ultrasound SMU-Net). These are correctness and reproducibility risks, not circularity. No equation or claim reduces to its own input; therefore the circularity score is 0.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

The method introduces no new physical entities. It relies on standard supervised training plus several hand-set hyperparameters and domain assumptions about BraTS data and fair baseline comparison. The most important free parameters are the loss weights and thresholds, which are not reported or sensitivity-tested.

free parameters (4)
  • Loss weights lambda_1, lambda_2, lambda_3, theta, lambda = not reported
    Balance the MSE distillation, BBDM, and LGRM losses in Eqs. (4) and (20). No values or sensitivity analysis are given, yet they control the final result.
  • Graph adjacency threshold tau = 0.8
    Threshold in Eq. (10) for binarizing the cross-graph cosine similarity matrix. Hand-set without sensitivity analysis.
  • Lesion-presence inference threshold = 0.5
    Hard gate in Eq. (16) that zeroes class probability volumes when predicted presence is at or below 0.5. Hand-set, not validated.
  • Number of graph anchors K and bandwidth vectors = not reported
    GraphNet3D uses K learnable anchors and channel-wise bandwidths in Eq. (6). K is not specified, and the anchors are trained on the target datasets.
axioms (5)
  • domain assumption BraTS annotations are a correct gold standard for tumor subregions.
    All Dice, IoU, HD95, and Sensitivity numbers are computed against these labels; the paper does not independently verify them.
  • domain assumption Zero-filling missing modalities plus 15 random combinations simulates real clinical missing-modality conditions.
    Training and evaluation rely on this simulation; no external dataset with naturally missing modalities is used.
  • domain assumption A teacher trained on full-modality data provides a reliable distillation target for incomplete student inputs.
    The entire BBDM design assumes that teacher knowledge transfers and improves the student; there is no independent support for this transferability beyond the paper's own results.
  • domain assumption The 20% test split of each dataset is representative and no distribution shift exists between train and test.
    A single random 80/20 split is used without cross-validation, so claims of generalization depend on this assumption.
  • domain assumption The 15 competing methods were implemented and tuned comparably.
    No baseline training or tuning protocol is documented, so the comparison assumes equal effort and fair hyperparameter selection.

pith-pipeline@v1.3.0-alltime-deepseek · 49943 in / 12585 out tokens · 95460 ms · 2026-08-04T16:12:15.444823+00:00 · methodology

0 comments
read the original abstract

Accurate brain tumor segmentation is essential for preoperative evaluation and personalized treatment. Multi-modal MRI is widely used due to its ability to capture complementary tumor features across different sequences. However, in clinical practice, missing modalities are common, limiting the robustness and generalizability of existing deep learning methods that rely on complete inputs, especially under non-dominant modality combinations. To address this, we propose AdaMM, a multi-modal brain tumor segmentation framework tailored for missing-modality scenarios, centered on knowledge distillation and composed of three synergistic modules. The Graph-guided Adaptive Refinement Module explicitly models semantic associations between generalizable and modality-specific features, enhancing adaptability to modality absence. The Bi-Bottleneck Distillation Module transfers structural and textural knowledge from teacher to student models via global style matching and adversarial feature alignment. The Lesion-Presence-Guided Reliability Module predicts prior probabilities of lesion types through an auxiliary classification task, effectively suppressing false positives under incomplete inputs. Extensive experiments on the Pretreat-MetsToBrain-Masks and BraTS 2018, 2024 datasets demonstrate that AdaMM consistently outperforms existing methods, exhibiting superior segmentation accuracy and robustness, particularly in single-modality and weak-modality configurations. In addition, we conduct a systematic evaluation of six categories of missing-modality strategies, supporting the superiority of knowledge distillation and offering practical guidance for method selection and future research. Our source code is available at https://github.com/Quanato607/AdaMM.

Figures

Figures reproduced from arXiv: 2509.15017 by Changmiao Wang, Feiwei Qin, Guanyu Zhou, Qiyuan Tian, Shenghao Zhu, Shuo Jiang, Weihong Chen, Yifei Chen, Yuanhan Wang.

Figure 1
Figure 1. Figure 1: (a) Data Generation: an external generator synthesizes absent modali￾ties, creating a full four-channel input for the segmentation model. (b) Feature Gen￾eration: the network learns to hallucinate modality-specific features internally when inputs are missing. (c) Sample Retrieval: the pipeline retrieves training cases from modality-matched cohorts to substitute for absent scans before segmentation. 2.1.1. … view at source ↗
Figure 2
Figure 2. Figure 2: (d) Robustness Enhancement: the network is trained with random modality dropout so it can segment directly from whatever scans are available. (e) Multi-task Learning: an auxiliary decoder reconstructs the absent modalities (red dashed arrows) while the main branch outputs the segmentation mask. (f) Knowledge Distillation: a full-modality teacher guides a partial-modality student through feature and predict… view at source ↗
Figure 3
Figure 3. Figure 3: Framework overview. (A) Missing-modality Sampling: Generates 15 MRI modality combinations and leverages an Adapter Bank to compensate for absent inputs. (B) Knowledge-distillation Training: Incorporate BBDM, GARM, and LGRM, with GARM applied exclusively to the student branch. 3.1. Bi-Bottleneck Distillation Module We perform Bi-Bottleneck Distillation Module on the GARM before and after the student model. … view at source ↗
Figure 4
Figure 4. Figure 4: Graph-guided Adaptive Refinement Module. (Stage 0) Graph–node Alignment, which establishes consistent correspondence between nodes across modality￾specific and shared graphs. (Stage 1) Cross-graph Enhancement, which facilitates mutual information exchange to enrich node representations. (Stage 2) Graph Refinement & Feature Reprojection, which progressively refines node embeddings and projects them back to … view at source ↗
Figure 5
Figure 5. Figure 5: Cosine similarity heatmaps of adapter parameters. Each four-digit code denotes available modalities in the order [T1, T1Gd, T2, FLAIR] (1=present, 0=ab￾sent). (a–c) 1st–3rd encoder adapters, (d–f) 1st–3rd decoder adapters. Deeper adapters show higher inter-configuration similarity, indicating convergence toward consistent full￾modality feature approximation, while retaining clear modality-combination speci… view at source ↗
Figure 6
Figure 6. Figure 6 [PITH_FULL_IMAGE:figures/full_fig_p017_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Scatter plots showing the relationship between model performance and parame [PITH_FULL_IMAGE:figures/full_fig_p020_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative comparation study on the BraTS 2024 dataset. Visual￾ization of a randomly selected sample from the BraTS 2024 dataset, showcasing segmen￾tation results under different modality-missing combinations across three distinct axial views. The corresponding multimodal MRI sequences and corresponding ground truth are presented, along with outputs of models. Dice and HD95 metrics are displayed for each … view at source ↗
Figure 9
Figure 9. Figure 9: Qualitative ablation study on the BraTS 2024 dataset. In￾put images and ground-truth masks are shown for four modality configurations (T1, T1+T1Gd+T2+FLAIR, T1Gd+T2, and T1+T1Gd+T2). From left to right, the remain￾ing columns present the segmentation outputs and corresponding prediction heatmaps produced by models without BBDM, without GARM, without LGRM, and with the com￾plete model. Dice scores for WT, T… view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

39 extracted references · 6 linked inside Pith

  1. [1]

    Havaei, A

    M. Havaei, A. Davy, D. Warde-Farley, A. Biard, A. Courville, Y. Bengio, C. Pal, P.-M. Jodoin, H. Larochelle, Brain tumor segmentation with deep neural networks, Medical Image Analysis 35 (2017) 18–31

    2017

  2. [2]

    Verhaert, P

    D. Verhaert, P. Thavendiranathan, S. Giri, G. Mihai, S. Rajagopalan, O. P. Simonetti, S. V. Raman, Direct T2 quantification of myocardial edema in acute ischemic injury, JACC: Cardiovascular Imaging 4 (2011) 269–278

    2011

  3. [3]

    J. N. Sarkaria, L. S. Hu, I. F. Parney, D. H. Pafundi, D. H. Brinkmann, N. N. Laack, C. Giannini, T. C. Burns, S. H. Kizilbash, J. K. Laramy, et al., Is the blood–brain barrier really disrupted in all glioblastomas? A critical assessment of existing clinical data, Neuro-oncology 20 (2018) 184–191

    2018

  4. [4]

    Bauer, R

    S. Bauer, R. Wiest, L.-P. Nolte, M. Reyes, A survey of MRI-based medical image analysis for brain tumor studies, Physics in Medicine & Biology 58 (2013) R97

    2013

  5. [5]

    C. He, Y. Shen, C. Fang, F. Xiao, L. Tang, Y. Zhang, W. Zuo, Z. Guo, X. Li, Diffusion models in low-level vision: A survey, IEEE Transactions on Pattern Analysis and Machine Intelligence 47 (2025) 4630–4651. 34

    2025

  6. [6]

    C. Lu, J. Zhang, D. Zhang, L. Mou, J. Yuan, K. Xia, Z. Guo, J. Zhang, Fine-Grained hierarchical progressive modal-aware network for brain tu- mor segmentation, IEEE Journal of Biomedical and Health Informatics (2025) 1–13

    2025

  7. [7]

    S. Wang, Y. Lin, Y. Wu, B. Du, Toward real ultra image segmenta- tion: Leveraging surrounding context to cultivate general segmentation model, in: Advances in Neural Information Processing Systems, vol- ume 37, 2024, pp. 129227–129249

    2024

  8. [8]

    Y. Chen, B. Zou, Z. Guo, Y. Huang, Y. Huang, F. Qin, Q. Li, C. Wang, SCUNet++: Swin-UNet and CNN bottleneck hybrid architecture with multi-fusion dense skip connection for pulmonary embolism CT image segmentation, in: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2024, pp. 7759–7767

    2024

  9. [9]

    Zhang, Y

    C. Zhang, Y. Chen, Z. Fan, Y. Huang, W. Weng, R. Ge, D. Zeng, C. Wang, TC-DiffRecon: Texture coordination MRI reconstruction method based on diffusion model and modified MF-UNet method, in: IEEE International Symposium on Biomedical Imaging, 2024, pp. 1–5

    2024

  10. [10]

    Y. Chen, S. Zhu, Z. Fang, C. Liu, B. Zou, L. Qiu, Y. Wang, S. Chang, F. Jia, F. Qin, J. Fan, Y. Peng, C. Wang, Toward robust early detection of alzheimer’s disease via an integrated multimodal learning approach, in: 2025 IEEE International Conference on Acoustics, Speech and Signal Processing, 2025, pp. 1–5

    2025

  11. [11]

    H.Zhou, F.Zhou, C.Zhao, Y.Xu, L.Luo, H.Chen, Multimodaldatain- tegration for precision oncology: Challenges and future directions, arXiv preprint arXiv:2406.19611 (2024)

    Pith/arXiv arXiv 2024

  12. [12]

    Goodfellow, J

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

    2020

  13. [13]

    D. P. Kingma, M. Welling, Auto-Encoding variational bayes, CoRR abs/1312.6114 (2013)

    Pith/arXiv arXiv 2013

  14. [14]

    Bakas, M

    S. Bakas, M. Reyes, A. Jakab, S. Bauer, M. Rempfler, A. Crimi, R. T. Shinohara, C. Berger, S. M. Ha, M. Rozycki, et al., Identifying the best 35 machine learning algorithms for brain tumor segmentation, progression assessment, and overall survival prediction in the BRATS challenge, arXiv preprint arXiv:1811.02629 (2018)

    Pith/arXiv arXiv 2018

  15. [15]

    M. C. de Verdier, R. Saluja, L. Gagnon, D. LaBella, U. Baid, N. H. Tahon, M. Foltyn-Dumitru, J. Zhang, M. Alafif, S. Baig, et al., The 2024 brain tumor segmentation (BraTS) challenge: Glioma segmentation on post-treatment MRI, arXiv preprint arXiv:2405.18368 (2024)

    Pith/arXiv arXiv 2024

  16. [16]

    Karras, M

    T. Karras, M. Aittala, J. Hellsten, S. Laine, J. Lehtinen, T. Aila, Train- ing generative adversarial networks with limited data, Advances in Neu- ral Information Processing Systems 33 (2020) 12104–12114

    2020

  17. [17]

    H. Wang, C. Ma, J. Zhang, Y. Zhang, J. Avery, L. Hull, G. Carneiro, Learnable cross-modal knowledge distillation for multi-modal learning with missing modality, in: International Conference on Medical Image Computing and Computer-Assisted Intervention, 2023, pp. 216–226

    2023

  18. [18]

    H. Wang, Y. Chen, C. Ma, J. Avery, L. Hull, G. Carneiro, Multi-modal learning with missing modality via shared-specific feature modelling, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 15878–15887

    2023

  19. [19]

    H. Ting, M. Liu, Multimodal Transformer of incomplete MRI data for brain tumor segmentation, IEEE Journal of Biomedical and Health Informatics 28 (2023) 89–99

    2023

  20. [20]

    Z.Yu, X.Han, S.Zhang, J.Feng, T.Peng, X.-Y.Zhang, MouseGAN++: unsupervised disentanglement and contrastive representation for mul- tiple MRI modalities synthesis and structural segmentation of mouse brain, IEEE Transactions on Medical Imaging 42 (2022) 1197–1209

    2022

  21. [21]

    H. Yang, J. Sun, Z. Xu, Learning unified hyper-network for multi-modal MR image synthesis and tumor segmentation with missing modalities, IEEE Transactions on Medical Imaging 42 (2023) 3678–3689

    2023

  22. [22]

    Z. Li, Y. Zhang, H. Li, Y. Chai, Y. Yang, Deformation-aware and reconstruction-driven multimodal representation learning for brain tu- morsegmentationwithmissingmodalities, BiomedicalSignalProcessing and Control 91 (2024) 106012. 36

    2024

  23. [23]

    Q. Fan, H. Yuan, H. Zuo, R. Liu, G. Gao, Leveraging retrieval augment approach for multimodal emotion recognition under missing modalities, arXiv preprint arXiv:2410.02804 (2024)

    Pith/arXiv arXiv 2024

  24. [24]

    Zhang, X

    C. Zhang, X. Chu, L. Ma, Y. Zhu, Y. Wang, J. Wang, J. Zhao, M3Care: Learning with missing modalities in multimodal healthcare data, in: Proceedings of the 28th ACM SIGKDD conference on knowledge dis- covery and data mining, 2022, pp. 2418–2428

    2022

  25. [25]

    Y. Sun, Z. Liu, Q. Z. Sheng, D. Chu, J. Yu, H. Sun, Similar modality completion-based multimodal sentiment analysis under uncertain miss- ing modalities, Information Fusion 110 (2024) 102454

    2024

  26. [26]

    X. Li, J. Jin, Y. Zhou, Y. Zhang, P. Zhang, Y. Zhu, Z. Dou, From matching to generation: A survey on generative information retrieval, ACM Transactions on Information Systems 43 (2025) 1–62

    2025

  27. [27]

    T. Wang, F. Li, L. Zhu, J. Li, Z. Zhang, H. T. Shen, Cross-Modal retrieval: A systematic review of methods and future directions, Pro- ceedings of the IEEE 112 (2024) 1716–1754

    2024

  28. [28]

    Y. Ding, X. Yu, Y. Yang, RFNet: Region-aware fusion network for incomplete multi-modal brain tumor segmentation, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 3975–3984

    2021

  29. [29]

    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: International Conference on Medical Image Computing and Computer-Assisted Inter- vention, 2022, pp. 107–117

    2022

  30. [30]

    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 Journal of Biomedical and Health Informatics 28 (2023) 379–390

    2023

  31. [31]

    C. Chen, Q. Dou, Y. Jin, H. Chen, J. Qin, P.-A. Heng, Robust multi- modal brain tumor segmentation via feature disentanglement and gated fusion, in: International Conference on Medical Image Computing and Computer-Assisted Intervention, 2019, pp. 447–456. 37

    2019

  32. [32]

    Jeong, H

    S. Jeong, H. Cho, J. Kwon, H. Park, Region-of-interest attentive hetero- modalvariationalencoder-decoderforsegmentationwithmissingmodal- ities, in: ProceedingsoftheAsianConferenceonComputerVision, 2022, pp. 3707–3723

    2022

  33. [33]

    S. Zhu, Y. Chen, S. Jiang, W. Chen, C. Liu, Y. Wang, X. Chen, Y. Ke, F. Qin, C. Wang, et al., XLSTM-HVED: Cross-modal brain tumor segmentation and MRI reconstruction method using vision XLSTM and heteromodal variational encoder-decoder, in: International Symposium on Biomedical Imaging, 2025, pp. 1–5

    2025

  34. [34]

    Y. Wang, Y. Zhang, Y. Liu, Z. Lin, J. Tian, C. Zhong, Z. Shi, J. Fan, Z. He, Acn: Adversarial co-training network for brain tumor segmenta- tion with missing modalities, in: International Conference on Medical Image Computing and Computer-Assisted Intervention, 2021, pp. 410– 420

    2021

  35. [35]

    Z. Ning, S. Zhong, Q. Feng, W. Chen, Y. Zhang, SMU-Net: Saliency- guided morphology-aware U-Net for breast lesion segmentation in ultra- sound image, IEEE Transactions on Medical Imaging 41 (2021) 476–490

    2021

  36. [36]

    S. Zhu, Y. Chen, W. Chen, Y. Wang, C. Liu, S. Jiang, F. Qin, C. Wang, Bridging the gap in missing modalities: Leveraging knowledge distilla- tion and style matching for brain tumor segmentation, arXiv preprint arXiv:2507.22626 (2025)

    Pith/arXiv arXiv 2025

  37. [37]

    D. Li, B. Yang, W. Zhan, X. He, Multi-category graph reasoning for multi-modal brain tumor segmentation, in: International Conference on Medical Image Computing and Computer-Assisted Intervention, 2024, pp. 445–455

    2024

  38. [38]

    Veličković, G

    P. Veličković, G. Cucurull, A. Casanova, A. Romero, P. Liò, Y. Ben- gio, Graph attention networks, International Conference on Learning Representations (2018)

    2018

  39. [39]

    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 Transactions on Medical Imaging 34 (2014) 1993–2024. 38

    2014