Pith. sign in

REVIEW 6 major objections 4 minor 57 references

F3-Net: Foundation Model for Full Abnormality Segmentation of Medical Images with Flexible Input Modality Requirement

T0 review · 6 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read F3-Net is a single segmentation network that covers multiple brain pathologies in one forward pass and is designed to tolerate missing MRI sequences through zero-filled inputs.

desk verdict Central claim of missing-modality robustness is never tested because no complete-input condition exists in any experiment; the paper stacks known components and the abstract overstates what is shown. read the letter →

arxiv 2507.08460 v1 pith:KJOW6OCM submitted 2025-07-11 cs.CV

Seyedeh Sahar Taheri Otaghsara , Reza Rahmanzadeh This is my paper
classification cs.CV
keywords medicalimagesegmentationfoundationmodelmissingMRImodalityzero-imagesynthesismulti-pathologybraintumorstrokelesionwhitematterhyperintensity
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

F3-Net claims that a single segmentation network can cover several brain pathologies—glioma, metastasis, ischemic stroke, and white matter hyperintensity—in one forward pass, without retraining per disease, and can accept whatever subset of six MRI sequences is available. The paper's mechanism is a multi-encoder nnU-Net with one encoder per modality and a shared decoder, trained on public datasets whose missing sequences are replaced by zero-filled volumes; during training the deepest encoder features from any zero-image modality are forcibly zeroed so the model learns from present channels. The authors build whole-pathology masks by merging each dataset's existing labels with WMH masks produced by a separately trained nnU-Net and manually corrected, then compare F3-Net both against these merged labels and against disease-specific models. If the claims hold, one deployment-ready model could replace separate pipelines for tumor and stroke segmentation and tolerate incomplete clinical MRI protocols. Reported average Dice scores reach 94.31% for BraTS-GLI 2024, 82.07% for BraTS-MET 2024, 94.12% for BraTS 2021, and 79.92% for ISLES 2022 on whole-pathology segmentation, with lower stroke-only performance than dedicated nnU-Net and SegResNet.

What carries the argument

F3-Net's central object is a multi-encoder nnU-Net variant: six separate encoders, one for each of T1, T1-Gd, T2, FLAIR, DWI, and ADC, feed a common bottleneck and a shared decoder. Missing sequences are supplied as zero-filled volumes, and the paper adds a training-time masking rule that forcibly zeroes the deepest encoder features of any zero-image modality before fusion, so the network cannot rely on empty channels. The other load-bearing component is the 'Pathoseg' label construction: a vanilla nnU-Net trained on WMH data produces FLAIR-space WMH masks that are manually corrected, co-registered, and merged with each dataset's existing pathology masks to yield binary whole-pathology targets. This machinery is what lets one model train across BraTS and ISLES with variable modality availability and output a comprehensive lesion map.

What would settle it

Take a held-out subset of BraTS or ISLES cases, run F3-Net with complete inputs, then rerun it with one or more sequences removed (for instance, dropping FLAIR or DWI) and compare average Dice. If the missing-sequence runs drop far below the complete-input scores, or match a model trained without the zero-image masking, the claimed robustness to absent modalities is not supported.

Watch

Extended reading notes

Core claim

The paper's central claim is that F3-Net, a unified architecture with modality-specific encoders and a shared decoder, can segment all visible brain pathology across tumor, metastasis, stroke, and white matter lesion cases in a single pass, while remaining accurate when MRI sequences are missing. Missing modalities are encoded as zero images, and at the deepest encoder stage features from these synthetic channels are set to zero so they do not affect the shared representation or backpropagation. Evaluated on whole-pathology masks that merge main disease labels with co-registered WMH segmentations, the model reports average Dice coefficients of 94.31% for BraTS-GLI 2024, 82.07% for BraTS-MET 2024, 94.12% for BraTS 2021, and 79.92% for ISLES 2022. On main-pathology comparisons it reports higher Dice than vanilla nnU-Net, Swin UNETR, and MedNeXt for the three BraTS settings, while on infarct-only ISLES 2022 it reports 77.28% versus 81.87% for nnU-Net and 82.23% for SegResNet.

Load-bearing premise

The load-bearing premise is that zero-filled missing MRI sequences, with the deepest encoder features of those sequences forcibly zeroed during training, preserve segmentation accuracy when sequences are actually absent—a premise the paper never tests, since all reported evaluations appear to use complete six-sequence inputs.

Editorial extensions

If this is right

  • A single F3-Net forward pass can return a unified lesion map covering tumor, metastasis, stroke, and white matter hyperintensity, removing the need for separate disease-specific segmenters for these conditions.
  • Clinical users can feed any available subset of the six MRI sequences without retraining, because absent sequences are zero-filled and masked during training.
  • For glioma and metastasis segmentation, F3-Net reports higher average Dice than vanilla nnU-Net, Swin UNETR, and MedNeXt on BraTS 2021 and BraTS 2024 main-pathology labels.
  • On stroke-only ISLES 2022, F3-Net remains competitive but does not beat dedicated models, reporting 77.28% average Dice versus 81.87% for nnU-Net and 82.23% for SegResNet.
  • Whole-pathology masks (Pathoseg) provide a benchmark setup in which multi-pathology segmentation can be measured directly, not just per-disease.

Reading between the lines

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

  • The paper asserts missing-modality robustness at the architectural level but reports no experiment with sequences actually removed; a direct ablation would show whether zero-fill plus deepest-feature masking preserves Dice as claimed.
  • Part of the reported performance margin may stem from how the whole-pathology masks were built: WMH labels came from a separately trained nnU-Net with manual correction, so gains on those labels partly reflect the quality of that auxiliary pipeline rather than F3-Net alone.
  • The fixed six-channel input with zero-fill could generalize to other modality sets, for example two-sequence stroke protocols, but the paper does not demonstrate that extension.
  • Because masking is applied only at the deepest encoder layer, intermediate features from zero images still propagate; ablating the masking depth would clarify whether the mechanism is necessary or incidental.
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

6 major / 4 minor

Summary. F3-Net proposes a multi-encoder nnU-Net that segments multiple brain pathologies (glioma, metastasis, stroke, white matter hyperintensity) in a single forward pass. The model expects six MRI modalities (T1, T1-Gd, T2, FLAIR, DWI, ADC) and handles missing sequences by filling absent channels with zero-valued images. A shared decoder is used after per-modality encoders, and features originating from zero-image channels are forcibly zeroed at the deepest encoder stage during training. The authors report Dice Similarity Coefficients on BraTS 2021, BraTS 2024 (glioma and metastasis), and ISLES 2022, for both 'whole pathology' masks (which include WMH) and 'main pathology' masks. The abstract claims robust performance under missing MRI sequences, generalizability across pathologies and domains, and superiority over CNN- and transformer-based models.

Significance. If the claims are validated, F3-Net would be a practically valuable step toward a single model that segments diverse brain lesions and tolerates incomplete acquisitions. The paper has several strengths: it addresses a clinically relevant problem, evaluates on multiple public benchmarks, uses strong baselines such as nnU-Net and Swin UNETR, and proposes a simple zero-image strategy that avoids explicit synthesis networks. The architectural description is clear, and the idea of separate modality encoders with a shared decoder is reasonable. However, the central advertised capability—robustness to missing MRI sequences—is never actually tested. All datasets used supply fewer than the six expected modalities, so every forward pass already contains zero-filled channels, and no comparison with a complete-input condition or an ablation removing an available modality is provided. The evaluation split is not stated, making the reported numbers difficult to interpret, and the 'whole pathology' ground truth is generated with the authors' own nnU-Net and manual correction, raising independence concerns.

major comments (6)
  1. [Section 5, Tables 1–3] The central claim of robustness to missing MRI sequences is not tested anywhere in Section 5. All datasets provide fewer than the six expected channels (BraTS provides four, ISLES provides three), so every forward pass already contains zero-filled channels; there is no experiment that compares the same model on complete versus incomplete input, and no ablation that removes an available modality. The abstract's sentence that F3-Net 'maintains robust performance even in the presence of missing MRI sequences' is therefore unsupported by the reported results.
  2. [Section 4, Training and Implementation Details] The training protocol does not describe any randomization of which modalities are available. Because all BraTS cases lack DWI and ADC and all ISLES cases lack T1, T1-Gd, and T2, the model only sees two fixed missing-modality patterns during training; it has never been trained with, for example, missing FLAIR but present DWI. Without modality dropout or synthetic missing-pattern augmentation, the claimed flexibility to arbitrary missing sequences cannot be learned, and the phrase 'flexible synthetic modality training' in the abstract is not supported.
  3. [Section 5, Tables 1–3] The manuscript does not state which data split is used for evaluation. For BraTS 2021, BraTS 2024, and ISLES 2022, official training, validation, and test partitions exist, but the paper does not say whether the reported DSCs come from a held-out set or from the training data. This is essential for interpreting all comparisons; without it, the numbers cannot be assessed.
  4. [Table 3 and Abstract] Table 3 shows F3-Net (77.28) below nnU-Net (81.87) and SegResNet (82.23) on ISLES 2022 main-pathology segmentation, yet the abstract and introduction state that F3-Net 'outperforms CNN-based and transformer-based models' without qualification. The claim should be restricted to the datasets and settings where it holds, and the ISLES result should be discussed as a known failure mode; no statistical significance or confidence intervals are provided for any of the comparisons.
  5. [Section 3.6] The 'whole pathology' ground truth (Pathoseg) is generated by a vanilla nnU-Net trained by the authors on WMH data, with outputs manually corrected by one neuroradiologist (RR). Because this same label-generation pipeline is used for the whole-pathology evaluation, the results are not independent of the authors' own models; the manuscript should clarify whether the same corrected WMH masks are used in training and testing, and discuss the potential bias this introduces.
  6. [Section 4] The zero-image masking is applied only at the deepest encoder stage before the bottleneck. Earlier encoder stages and skip connections from the zero channels are left unmasked, so those channels can still propagate information to the decoder. The statement that zero inputs 'do not contribute to the learned representations' is therefore inaccurate, and the mechanism should be specified at all stages or justified as sufficient.
minor comments (4)
  1. [Table 1] Table 1 contains a row 'BraTS 2024' with an average DSC of 93.95 in addition to separate rows for BraTS-GLI 2024 and BraTS-MET 2024; the paper should state how this combined average is computed.
  2. [Sections 1 and 4] The phrase 'flexible synthetic modality training' is used in the abstract and introduction but is not defined in Sections 3–4; the training section should describe the exact procedure for creating synthetic missing-modality examples.
  3. [Throughout] There are several typographical errors, including 'T able 1' on page 8 and 'synthesised' versus 'synthesized' inconsistency; a careful proofread is needed.
  4. [Implementation Details] The paper does not mention code or model availability; providing the trained model and inference code would strengthen reproducibility.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation found; the central claims are empirical and under-supported in places, not reduced to their inputs.

full rationale

The paper contains no equation-level reduction of a predicted quantity to a fitted input. Its central results are benchmark DSCs obtained by supervised training and inference, not derived from a self-referential formula. The only self-citation is reference [6] (the authors' prior multi-encoder nnU-Net work), which is cited as related work and not used to justify a forbidden alternative or to import a uniqueness theorem; it is not load-bearing. The whole-pathology ground truth in Section 3.6 is created by training a vanilla nnU-Net on WMH data, manually correcting its outputs by a neuroradiologist, co-registering the masks, and merging them with existing dataset labels. This is a data-generation dependency for the WMH component of the Table 1 scores, but it is not circular by construction: manual correction intervenes, and F3-Net's outputs are not equal to the label-generator's outputs by definition. The abstract's missing-modality robustness claim is not circular either, but it is under-supported: because no evaluated dataset supplies all six expected sequences (BraTS supplies four, ISLES three), every forward pass already contains zero-filled channels, and the paper reports no complete-input baseline and no ablation that removes an available modality, despite the Introduction promising comparison 'under varying modality configurations'. That is a reporting gap and a correctness risk, not a circularity. A score of 1 reflects the minor self-citation and the model-assisted label loop, while recognizing that the main derivation chain is independent of these dependencies.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

No new physical entities are introduced. The main burden is the label-generation loop: F3-Net is trained and evaluated on masks partly produced by another model (the WMH nnU-Net) and corrected by an author. This makes the reported DSCs depend on unvalidated pseudo-labels.

free parameters (3)
  • Loss weighting λ1, λ2 = not reported
    Equation 1 defines the loss as a weighted sum of Dice and cross-entropy, but the numerical values of λ1 and λ2 are not given, so the training objective is underspecified.
  • Modality dropout schedule = not reported
    The paper states training uses zero-filled synthetic modalities but never gives the probability or schedule for masking modalities, which is essential for learning missing-modality robustness.
  • Zero-masking layer = deepest encoder layer before bottleneck
    The choice of which layer's features to zero for missing modalities is a design decision without ablation or justification.
assumptions (4)
  • domain assumption BraTS 2021, BraTS 2024, and ISLES 2022 ground truth labels are accurate.
    The evaluation uses these challenge labels as reference without independent verification.
  • domain assumption The WMH masks produced by a vanilla nnU-Net and manually corrected by one neuroradiologist are valid segmentation ground truth.
    Section 3.6 describes this pipeline; no inter-rater reliability or external validation is reported.
  • ad hoc to paper Zero-filled images with feature masking encode missing modalities without harming segmentation.
    This is the central mechanism for handling missing inputs, but no experiment tests it; the paper assumes it works.
  • domain assumption Evaluation is performed on a held-out set separate from training.
    The manuscript never states how cases were split, so there is no evidence against training-set evaluation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of F3-Net: Foundation Model for Full Abnormality Segmentation of Medical Images with Flexible Input Modality Requirement." pith.science (2026). https://pith.science/paper/KJOW6OCM

@misc{pith2026250708460,
  author       = {Pith},
  title        = {Pith review of: F3-Net: Foundation Model for Full Abnormality Segmentation of Medical Images with Flexible Input Modality Requirement},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KJOW6OCM}},
  note         = {Machine review of arXiv:2507.08460}
}
read the original abstract

F3-Net is a foundation model designed to overcome persistent challenges in clinical medical image segmentation, including reliance on complete multimodal inputs, limited generalizability, and narrow task specificity. Through flexible synthetic modality training, F3-Net maintains robust performance even in the presence of missing MRI sequences, leveraging a zero-image strategy to substitute absent modalities without relying on explicit synthesis networks, thereby enhancing real-world applicability. Its unified architecture supports multi-pathology segmentation across glioma, metastasis, stroke, and white matter lesions without retraining, outperforming CNN-based and transformer-based models that typically require disease-specific fine-tuning. Evaluated on diverse datasets such as BraTS 2021, BraTS 2024, and ISLES 2022, F3-Net demonstrates strong resilience to domain shifts and clinical heterogeneity. On the whole pathology dataset, F3-Net achieves average Dice Similarity Coefficients (DSCs) of 0.94 for BraTS-GLI 2024, 0.82 for BraTS-MET 2024, 0.94 for BraTS 2021, and 0.79 for ISLES 2022. This positions it as a versatile, scalable solution bridging the gap between deep learning research and practical clinical deployment.

Figures

Figures reproduced from arXiv: 2507.08460 by the authors.

Figure 1
Figure 1. Pipeline for generating comprehensive pathology segmentation masks (Pathoseg). [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Overview of the F3-Net architecture. Each MRI modality is processed by its own dedi￾cated encoder, enabling the extraction of features that are specifically adapted to the distinctive properties of each modality. At the bottleneck stage, the encoded features from all modalities are merged, blending their diverse information to create a unified and robust feature representation. This integrated repre￾sentation is sub… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

57 extracted references · 46 canonical work pages

  1. [1]

    Clinical impact of deep learning reconstruction in mri,

    S. Kiryu, H. Akai, K. Yasaka, T. Tajima, A. Kunimatsu, N. Yoshioka, M. Akahane, O. Abe, and K. Ohtomo, “Clinical impact of deep learning reconstruction in mri,” Radiographics, vol. 43, no. 6, p. e220133, 2023

    work page 2023

  2. [2]

    Deep learning for image en- hancement and correction in magnetic resonance imaging—state-of-the-art and challenges,

    Z. Chen, K. Pawar, M. Ekanayake, C. Pain, S. Zhong, and G. F. Egan, “Deep learning for image en- hancement and correction in magnetic resonance imaging—state-of-the-art and challenges,” Journal of Digital Imaging , vol. 36, no. 1, pp. 204–230, 2023

    work page 2023

  3. [3]

    Narrative review of generative adversarial networks in medical and molecular imaging,

    S. Koshino et al. , “Narrative review of generative adversarial networks in medical and molecular imaging,” Annals of Translational Medicine , 2021

    work page 2021

  4. [4]

    Foundation models for biomedical image segmen- tation: A survey,

    H. H. Lee, Y. Gu, T. Zhao, Y. Xu, J. Yang et al., “Foundation models for biomedical image segmen- tation: A survey,” arXiv preprint arXiv:2401.07654 , 2024

    arXiv 2024

  5. [5]

    On the challenges and perspectives of foundation models for medical image analysis,

    S. Zhang and D. Metaxas, “On the challenges and perspectives of foundation models for medical image analysis,” Medical Image Analysis, vol. 91, p. 102996, 2023

    work page 2023

  6. [6]

    Multi-encoder nnU-Net outperforms transformer models with self-supervised pretraining

    S. S. Otaghsara and R. Rahmanzadeh, “Multi-encoder nnu-net outperforms transformer models with self-supervised pretraining,” arXiv preprint arXiv:2504.03474 , 2025

    work page Pith review arXiv 2025

  7. [7]

    Medical sam 2: Segment medical images as video via segment anything model 2,

    J. Zhu, A. Hamdi, Y. Qi, Y. Jin, and J. Wu, “Medical sam 2: Segment medical images as video via segment anything model 2,” 2024. [Online]. Available: https://arxiv.org/abs/2408.00874

    arXiv 2024

  8. [8]

    Foundation AI Model for Medical Image Segmentation

    R. Bao, E. Darzi, S. He et al., “Foundation ai model for medical image segmentation,” arXiv preprint arXiv:2411.02745, 2024

    work page Pith review arXiv 2024

Show all 57 references
  1. [9]

    Generative adversarial networks (gan) powered fast magnetic resonance imaging – mini review, comparison, and perspectives,

    G. Yang et al. , “Generative adversarial networks (gan) powered fast magnetic resonance imaging – mini review, comparison, and perspectives,” arXiv preprint arXiv:2105.01800 , 2021

    2021 arXiv

  2. [10]

    Gans for medical image synthesis: An empirical study,

    Y. Skandarani, P.-M. Jodoin, and A. Lalande, “Gans for medical image synthesis: An empirical study,” arXiv preprint arXiv:2105.05318 , 2021

    2021 arXiv

  3. [11]

    Generative adversarial networks for brain images synthesis: A review,

    F. Shomal Zadeh et al. , “Generative adversarial networks for brain images synthesis: A review,” arXiv preprint arXiv:2305.15421 , 2023

    2023 arXiv

  4. [12]

    Generative adversarial networks: A primer for radiologists,

    RadioGraphics, “Generative adversarial networks: A primer for radiologists,” 2020, online resource

    2020

  5. [13]

    Vision-mae: A foundation model for medical image segmentation and classification,

    Z. Liu, A. Tieu, N. Patel et al. , “Vision-mae: A foundation model for medical image segmentation and classification,” arXiv preprint arXiv:2402.01034 , 2024

    2024 arXiv

  6. [14]

    Ensembles of multiple models and architectures for robust brain tumour segmentation,

    K. Kamnitsas et al. , “Ensembles of multiple models and architectures for robust brain tumour segmentation,” in International MICCAI Brainlesion Workshop . Springer, 2017, pp. 450–462

    2017

  7. [15]

    nnu-net: a self-configuring method for deep learning-based biomedical image segmentation,

    F. Isensee et al. , “nnu-net: a self-configuring method for deep learning-based biomedical image segmentation,” Nature Methods, vol. 18, no. 2, pp. 203–211, 2021

    2021

  8. [16]

    Unetr: Transformers for 3d medical image segmentation,

    H. Hatamizadeh et al., “Unetr: Transformers for 3d medical image segmentation,” in Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2022, pp. 574–584. 9

    2022

  9. [17]

    Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images,

    A. Hatamizadeh, V. Nath, Y. Tang, D. Yang, H. Roth, and D. Xu, “Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images,” 2022. [Online]. Available: https://arxiv.org/abs/2201.01266

    2022 arXiv

  10. [18]

    Mednext: Transformer-driven scaling of convnets for medical image segmentation,

    S. Roy et al. , “Mednext: Transformer-driven scaling of convnets for medical image segmentation,” arXiv preprint arXiv:2303.09975 , 2023

    2023 arXiv

  11. [19]

    Brats challenge 2023 & 2024 results summary,

    BraTS Challenge Organizers, “Brats challenge 2023 & 2024 results summary,” https://www.med. upenn.edu/sbia/brats2023.html, accessed April 2025

    2023

  12. [20]

    Visual language pretrained multiple instance zero-shot transfer for histopathology images,

    M. Y. Lu, B. Chen, A. Zhang, D. F. K. Williamson, R. J. Chen, T. Ding, L. P. Le, Y.-S. Chuang, and F. Mahmood, “Visual language pretrained multiple instance zero-shot transfer for histopathology images,” 2023. [Online]. Available: https://arxiv.org/abs/2306.07831

    2023 arXiv

  13. [21]

    Learning to exploit temporal structure for biomedical vision-language processing,

    S. Bannur, S. Hyland, Q. Liu, F. Perez-Garcia, M. Ilse, D. C. Castro, B. Boecking, H. Sharma, K. Bouzid, A. Thieme et al., “Learning to exploit temporal structure for biomedical vision-language processing,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Patte...

    2023

  14. [22]

    Segment anything,

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y. Lo et al., “Segment anything,” inProceedings of the IEEE/CVF international conference on computer vision , 2023, pp. 4015–4026

    2023

  15. [23]

    Segment anything in medical images,

    J. Ma, Y. He, F. Li, L. Han, C. You, and B. Wang, “Segment anything in medical images,” Nature Communications, vol. 15, no. 1, p. 654, 2024

    2024

  16. [24]

    Medlsam: Localize and segment anything model for 3d medical images,

    X. Z. K. L. S. Z. Wenhui Lei, Xu Wei, “Medlsam: Localize and segment anything model for 3d medical images,” arXiv preprint arXiv: , 2023

    2023

  17. [25]

    Sam-med2d: Segment anything model for medical image segmentation on 2d images,

    X. Zhang et al. , “Sam-med2d: Segment anything model for medical image segmentation on 2d images,” arXiv preprint arXiv:2307.02742 , 2023

    2023 arXiv

  18. [26]

    Sam- med3d: towards general-purpose segmentation models for volumetric medical images,

    H. Wang, S. Guo, J. Ye, Z. Deng, J. Cheng, T. Li, J. Chen, Y. Su, Z. Huang, Y. Shen et al. , “Sam- med3d: towards general-purpose segmentation models for volumetric medical images,” in European Conference on Computer Vision . Springer, 2025, pp. 51–67

    2025

  19. [27]

    Modality-agnostic medical image segmentation via unified representation learning,

    D. Karimi, S. Nir, and A. Gholipour, “Modality-agnostic medical image segmentation via unified representation learning,” Medical Image Analysis, vol. 78, p. 102389, 2022

    2022

  20. [28]

    Generalizable medical image segmentation via modality-invariant latent space,

    Y. Chen et al. , “Generalizable medical image segmentation via modality-invariant latent space,” in CVPR Workshops , 2023

    2023

  21. [29]

    Unified cross-modal pretraining for medical image understanding,

    H. Li et al. , “Unified cross-modal pretraining for medical image understanding,” arXiv preprint arXiv:2303.05677, 2023

    2023 arXiv

  22. [31]

    The first step for neuroimaging data analysis: Dicom to nifti conversion,

    X. Li, P. S. Morgan, J. Ashburner, J. Smith, and C. Rorden, “The first step for neuroimaging data analysis: Dicom to nifti conversion,” Journal of neuroscience methods , vol. 264, pp. 47–56, 2016

    2016

  23. [32]

    Ensembles of densely-connected cnns with label-uncertainty for brain tumor segmentation,

    R. McKinley, R. Meier, and R. Wiest, “Ensembles of densely-connected cnns with label-uncertainty for brain tumor segmentation,” in International MICCAI brainlesion workshop . Springer, 2018, pp. 456–465

    2018

  24. [33]

    Multi-scale 3d convolutional neural networks for lesion segmentation in brain mri,

    K. Kamnitsas, L. Chen, C. Ledig, D. Rueckert, B. Glocker et al. , “Multi-scale 3d convolutional neural networks for lesion segmentation in brain mri,” Ischemic stroke lesion segmentation , vol. 13, p. 46, 2015

    2015

  25. [34]

    Efficient multi-scale 3d cnn with fully connected crf for accurate brain lesion segmentation,

    K. Kamnitsas, C. Ledig, V. F. Newcombe, J. P. Simpson, A. D. Kane, D. K. Menon, D. Rueckert, and B. Glocker, “Efficient multi-scale 3d cnn with fully connected crf for accurate brain lesion segmentation,” Medical image analysis , vol. 36, pp. 61–78, 2017

    2017

  26. [35]

    Simultaneous truth and performance level estimation (staple): an algorithm for the validation of image segmentation,

    S. K. Warfield, K. H. Zou, and W. M. Wells, “Simultaneous truth and performance level estimation (staple): an algorithm for the validation of image segmentation,” IEEE transactions on medical imaging, vol. 23, no. 7, pp. 903–921, 2004. 10

    2004

  27. [36]

    The 2024 brain tumor segmentation (brats) challenge: Glioma segmentation on post-treatment mri,

    M. C. de Verdier, R. Saluja, L. Gagnon, D. LaBella, U. Baid, N. H. Tahon, M. Foltyn-Dumitru, J. Zhang, M. Alafif, S. Baig, K. Chang, G. D’Anna, L. Deptula, D. Gupta, M. A. Haider, A. Hussain, M. Iv, M. Kontzialis, P. Manning, F. Moodi, T. Nunes, A. Simon, N. Sollmann, D. Vu, M...

    2024 arXiv

  28. [37]

    Automated brain extraction of multisequence mri using artificial neural networks,

    F. Isensee, M. Schell, I. Pflueger, G. Brugnara, D. Bonekamp, U. Neuberger, A. Wick, H. Schlemmer, S. Heiland, W. Wick, M. Bendszus, K. H. Maier-Hein, and P. Kickingereder, “Automated brain extraction of multisequence mri using artificial neural networks,” Human Brain Mapping ...

    2019 doi

  29. [38]

    Isles 2022: A multi-center magnetic resonance imaging stroke lesion segmentation dataset,

    M. R. Hernandez Petzsche, E. de la Rosa, U. Hanning, R. Wiest, W. Valenzuela, M. Reyes, M. Meyer, S.-L. Liew, F. Kofler, I. Ezhov et al., “Isles 2022: A multi-center magnetic resonance imaging stroke lesion segmentation dataset,” Scientific data , vol. 9, no. 1, p. 762, 2022

    2022

  30. [39]

    Data of the White Matter Hyperintensity (WMH) Segmentation Challenge,

    H. Kuijf, M. Biesbroek, J. de Bresser, R. Heinen, C. Chen, W. van der Flier, Barkhof, M. Viergever, and G. J. Biessels, “Data of the White Matter Hyperintensity (WMH) Segmentation Challenge,”

  31. [40]

    Shifts 2.0: Extending the dataset of real distribu- tional shifts,

    A. Malinin, A. Athanasopoulos, M. Barakovic, M. B. Cuadra, M. J. Gales, C. Granziera, M. Graziani, N. Kartashev, K. Kyriakopoulos, P.-J. Lu et al., “Shifts 2.0: Extending the dataset of real distribu- tional shifts,” arXiv preprint arXiv:2206.15407 , 2022

    2022 arXiv

  32. [41]

    Brain tumor segmentation on mri with missing modalities,

    Y. Shen and M. Gao, “Brain tumor segmentation on mri with missing modalities,” in Information Processing in Medical Imaging (IPMI), ser. Lecture Notes in Computer Science, vol. 11492. Springer, 2019, pp. 417–428

    2019

  33. [42]

    Deep learning-based imaging data completion for improved brain disease diagnosis,

    H. Li, M. Habes, Y. Fan, and C. Davatzikos, “Deep learning-based imaging data completion for improved brain disease diagnosis,” in International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2014, pp. 305–312

    2014

  34. [43]

    Brain tumor segmentation with deep neural networks,

    M. Havaei, A. Davy, D. Warde-Farley, A. Biard, A. Courville, Y. Bengio, C. Pal, P.-M. Jodoin, and H. Larochelle, “Brain tumor segmentation with deep neural networks,” Medical image analysis, vol. 35, pp. 18–31, 2017

    2017

  35. [44]

    M3ae: Multimodal representation learning for brain tumor segmentation with missing modalities,

    H. Liu, D. Wei, D. Lu, J. Sun, L. Wang, and Y. Zheng, “M3ae: Multimodal representation learning for brain tumor segmentation with missing modalities,” arXiv preprint arXiv:2303.05302 , 2023

    2023 arXiv

  36. [45]

    Unified multi-modal image synthesis for missing modality imputation,

    Y. Zhang, C. Peng, Q. Wang, D. Song, K. Li, and S. K. Zhou, “Unified multi-modal image synthesis for missing modality imputation,” arXiv preprint arXiv:2304.05340 , 2023

    2023 arXiv

  37. [46]

    Mri scan synthesis methods based on clus- tering and pix2pix,

    G. Baldini, M. Schmidt, C. Z¨ aske, and L. L. Caldeira, “Mri scan synthesis methods based on clus- tering and pix2pix,” arXiv preprint arXiv:2312.05176 , 2023

    2023 arXiv

  38. [47]

    Multi-modal modality-masked diffusion network for brain mri synthesis with random modality missing,

    X. Meng, K. Sun, J. Xu, X. He, and D. Shen, “Multi-modal modality-masked diffusion network for brain mri synthesis with random modality missing,” IEEE Transactions on Medical Imaging, vol. 43, no. 7, pp. 2587–2598, 2024

    2024

  39. [48]

    Highest resolution in vivo human brain mri using prospective motion correction,

    D. Stucht, K. A. Danishad, P. Schulze, F. Godenschweger, M. Zaitsev, and O. Speck, “Highest resolution in vivo human brain mri using prospective motion correction,” PloS one , vol. 10, no. 7, p. e0133921, 2015

    2015

  40. [49]

    Swinbts: A method for 3d multimodal brain tumor segmentation using swin transformer,

    Y. Jiang, Y. Zhang, X. Lin, J. Dong, T. Cheng, and J. Liang, “Swinbts: A method for 3d multimodal brain tumor segmentation using swin transformer,” Brain sciences, vol. 12, no. 6, p. 797, 2022

    2022

  41. [50]

    Seresu-net for multimodal brain tumor segmentation,

    C. Yan, J. Ding, H. Zhang, K. Tong, B. Hua, and S. Shi, “Seresu-net for multimodal brain tumor segmentation,” IEEE Access, vol. 10, pp. 117 033–117 044, 2022. 11

    2022

  42. [51]

    Factorised spatial representation learning: Application in semi-supervised my- ocardial segmentation,

    M. Chartsias et al., “Factorised spatial representation learning: Application in semi-supervised my- ocardial segmentation,” in MICCAI, 2019

    2019

  43. [52]

    Self-supervised 3d patient modeling with multi-modal attentive fusion,

    M. Zheng, B. Planche, X. Gong, F. Yang, T. Chen, and Z. Wu, “Self-supervised 3d patient modeling with multi-modal attentive fusion,” 2024. [Online]. Available: https://arxiv.org/abs/2403.03217

    2024 arXiv

  44. [53]

    Synthseg: Segmentation of brain mri scans of any contrast and resolution without retraining,

    B. Billot, D. N. Greve, O. Puonti, A. Thielscher, K. Van Leemput, B. Fischl, A. V. Dalca, J. E. Iglesias et al. , “Synthseg: Segmentation of brain mri scans of any contrast and resolution without retraining,” Medical image analysis , vol. 86, p. 102789, 2023

    2023

  45. [54]

    Deep generative adversarial networks: applications in muscu- loskeletal imaging,

    Y. Shin, J. Yang, and Y. H. Lee, “Deep generative adversarial networks: applications in muscu- loskeletal imaging,” Radiology: Artificial Intelligence , vol. 3, no. 3, p. e200157, 2021

    2021

  46. [55]

    Medical image synthesis for data augmentation and anonymization using gans,

    H. Shin et al. , “Medical image synthesis for data augmentation and anonymization using gans,” in Springer Lecture Notes in Computer Science (LNCS) , 2018

    2018

  47. [56]

    The rsna-asnr-miccai brats 2021 benchmark on brain tumor segmentation and radiogenomic classification,

    U. Baid, S. Ghodasara, S. Mohan, M. Bilello, E. Calabrese, E. Colak, K. Farahani, J. Kalpathy-Cramer, F. C. Kitamura, S. Pati, L. M. Prevedello, J. D. Rudie, C. Sako, R. T. Shinohara, T. Bergquist, R. Chai, J. Eddy, J. Elliott, W. Reade, T. Schaffter, T. Yu, J. Zheng, A. W. Mo...

    2021 arXiv

  48. [57]

    Domain adaptation for medical image analysis: a survey,

    H. Guan and M. Liu, “Domain adaptation for medical image analysis: a survey,” IEEE Transactions on Biomedical Engineering, vol. 69, no. 3, pp. 1173–1185, 2021. 12

    2021

  49. [2022]

    Available: https://doi.org/10.34894/AECRSD

    [Online]. Available: https://doi.org/10.34894/AECRSD

    doi

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.