Pith. sign in

REVIEW 4 major objections 5 minor 25 references

The paper claims that masking only organ-rich subvolumes—defined by a mean Hounsfield-unit threshold of 0.1—during masked image modeling pretraining consistently improves 3D medical segmentation, reporting top Dice scores of 84.64% (BTCV),

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 22:01 UTC pith:TAHABPE6

load-bearing objection Simple HU-threshold masking prior for 3D MIM; plausible idea, but evidence is under-powered and the method description has a contradictory equation. the 4 major comments →

arxiv 2509.07534 v1 pith:TAHABPE6 submitted 2025-09-09 cs.CV cs.AI

HU-based Foreground Masking for 3D Medical Masked Image Modeling

Jin Lee , Vu Dang , Gwang-Hyun Yu , Anh Le , Zahid Rahman , Jin-Ho Jang , Heonzoo Lee , Kun-Yung Kim
show 2 more authors
Jin-Sul Kim Jin-Young Kim
This is my paper
classification cs.CV cs.AI
keywords masked image modeling3D medical image segmentationHounsfield unitsself-supervised learningCT pretrainingforeground maskingUNETRSwinUNETR
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.

Masked image modeling pretraining for 3D medical images normally masks random subvolumes, but much of a CT volume is air or fluid that carries no anatomical signal. The paper's claim is that the pretraining mask should instead be driven by the physical intensity calibration of CT: keep only 16×16×16 subvolumes whose mean Hounsfield unit value is at least 0.1, and train the model to reconstruct the full volume from that organ-rich input. Across five public benchmarks, this simple change consistently improves downstream segmentation Dice over random masking, local masking, and several self-supervised baselines, for both UNETR and SwinUNETR backbones. The result matters because it suggests a nearly free, domain-grounded way to make medical MIM pretraining more sample-efficient, and it works even when the pretraining is on CT and the downstream task is MRI brain-tumor segmentation.

Core claim

The central discovery is that an intensity-defined foreground mask is a better pretext curriculum for 3D medical MIM than random or structurally local masks. On 40 CT images, the paper observes that background (mean HU below 0.1) has lower entropy, complexity, and mutual information with labels than foreground, so it defines a masking operator that zeros out every 16^3 subvolume whose mean HU falls below λ=0.1 and keeps the rest. The masked input is fed to a transformer (UNETR or SwinUNETR) trained with an L1 reconstruction loss to recover the original volume. On downstream fine-tuning, the paper reports that this foreground-only pretraining outperforms all compared baselines on BTCV, Flare2

What carries the argument

The load-bearing object is the HU-based Foreground Masking operator defined over subvolumes: after partitioning a volume into non-overlapping 16×16×16 cubes, the mean Hounsfield-unit value of each cube is compared with a fixed threshold λ=0.1, and cubes below threshold are zeroed out of the masked input. This turns the physical HU scale (air at −1000, fluid near 0, dense tissue positive) into a binary curriculum: the pretext model sees only organ-bearing context and must reconstruct the discarded background and the retained foreground. The same operator, with the same threshold, is applied unchanged to four CT datasets and one MRI dataset, which is what makes the reported transfer and consis

Load-bearing premise

The load-bearing premise is that a single fixed threshold, λ=0.1 on the mean Hounsfield value of each 16^3 subvolume, cleanly separates informative organ tissue from uninformative background across all four CT datasets and the MRI dataset; if organ intensities fall below that threshold or a different normalization shifts intensities, the mask removes informative tissue and the reported gains should reverse.

What would settle it

Run the same pretraining pipeline on a CT dataset whose target organs have mean HU below 0.1—for example a low-contrast soft-tissue or delayed-contrast-phase study—or after rescaling intensities so organ subvolumes fall below λ. If the fixed threshold then performs no better than random masking, or if tuning λ per dataset materially changes Dice, the claim that a global HU threshold is the active ingredient fails.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Random masking is not a neutral choice for medical MIM: replacing it with HU-based foreground masking improves average Dice on all five tested datasets for both UNETR and SwinUNETR backbones.
  • The benefit survives a change of modality: models pretrained only on CT transfer to MRI brain-tumor segmentation (BraTS) and still beat all compared baselines.
  • A masking ratio of 0.6 is the best operating point in ablations, and the fixed threshold λ=0.1 is chosen from an analysis of 40 CT images rather than tuned per dataset.
  • On the SwinUNETR backbone, the method's scores are comparable to a recent volume-contrastive baseline (VoCo) while using only the reconstruction pretext task and no contrastive loss.
  • The improvement over training from scratch is 2.91 Dice points for UNETR and 2.06 points for SwinUNETR averaged over three datasets, showing the pretraining signal is not a wash.

Where Pith is reading between the lines

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

  • Pith inference: the fixed threshold λ=0.1 is the fragile part of the recipe; if a dataset's normalization or contrast phase puts organ subvolumes below the threshold, the mask will discard informative tissue, so an adaptive or per-dataset threshold is the natural next experiment.
  • Pith inference: the same logic should transfer to any imaging modality with a calibrated intensity scale—PET standardized uptake values, quantitative susceptibility maps, or dual-energy CT material densities—where foreground can be defined without labels.
  • Pith inference: because the mask selects what the reconstruction loss sees, weighting the L1 loss by the same foreground indicator (or masking ratio) might interact with the gain; the paper does not ablate loss weighting, so that is open.
  • Pith inference: the entropy/complexity/mutual-information analysis on Flare22 is on one dataset; a broader test across CT contrast phases and body regions would show whether the 0.1 threshold is a stable property of CT physics or an artifact of the analysis set.

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 / 5 minor

Summary. The paper proposes HU-based Foreground Masking for 3D medical Masked Image Modeling (MIM). The method partitions a CT volume into 16×16×16 subvolumes, computes the mean HU for each subvolume, and masks only those subvolumes whose mean HU exceeds a global threshold λ=0.1. The pretext task is standard MIM (L1 reconstruction of the masked volume), but the loss is restricted to the masked foreground subvolumes. The authors evaluate on five public datasets (BTCV, FLARE22, MM-WHS, AMOS22, BraTS) with UNETR and SwinUNETR backbones, comparing against random/local masking, MAE, SimMIM, and several medical pre-training baselines such as VoCo and SwinMM. They report consistent Dice improvements, e.g., BTCV 84.64% with SwinUNETR, and claim that domain-centric masking is important for medical image segmentation.

Significance. If the claimed improvements are robust, the paper offers a simple, domain-specific masking strategy that could be a useful baseline for 3D medical MIM. The approach is easy to implement and the authors release code, which supports reproducibility. The evaluation spans multiple CT datasets and one MRI dataset, covering both multi-organ and brain tumor segmentation. However, the current evidence is insufficient to establish the central claim of consistent improvement: the margins over strong baselines are small (0.32% over VoCo, 1.21% over SimMIM), the masking-ratio hyperparameter is selected on the same BTCV split used for a headline result, and no error bars or significance tests are provided. The transfer to MRI (BraTS) is also not justified given that HU is not defined for MRI.

major comments (4)
  1. [§3.3, Fig. 6b] The masking ratio 0.6 is selected by sweeping on BTCV (Fig. 6b), yet BTCV is also a headline evaluation dataset in §3.2. This selection on the evaluation split can inflate the reported 84.64% Dice. The margin over random (0.66%) and local (0.75%) masking in Fig. 6a is small enough that selection-on-validation could account for the difference. Please report results for all swept ratios on every dataset, or use a nested/leave-out validation scheme, and state whether the random/local baselines and SimMIM/VoCo received the same per-dataset hyperparameter tuning. Without this, the claim of consistent improvement is not established.
  2. [§3.2, Fig. 4] No error bars, repeated seeds, or significance tests are reported. The improvements over the strongest baselines are 1.21% over SimMIM (UNETR) and 0.32% over VoCo (SwinUNETR); the masking-strategy ablation shows 0.75% over local masking. These margins are within typical seed-to-seed variance for 3D segmentation with small validation sets (e.g., BTCV has only 6 validation scans). Please run at least 3 seeds, report mean±std, and perform paired significance tests (e.g., bootstrap or Wilcoxon over validation cases). Also report per-organ variance if available.
  3. [§2.2, Eq. (2)-(3); §4] The single global threshold λ=0.1 is fitted to "40 CT images" and is then applied to all CT datasets and to BraTS, which is MRI. The paper itself concedes in §4 that a single global threshold may not be optimal across subvolumes and datasets. For MRI, Hounsfield Units are not defined; the normalized intensities may not have the same semantics. Please (i) clarify the origin of the 40 CT images and their overlap with pretraining/downstream data, (ii) provide a sensitivity analysis of λ on each dataset (including BraTS) and show that the reported gains hold over a range of λ, and (iii) explain how the threshold transfers to MRI or restrict the claim to CT.
  4. [§2.1, Fig. 3] The ROI analysis uses the same foreground/background definition (HU < 0.1 as background) that the method is built on. This is therefore descriptive rather than an independent validation of the threshold. Please reframe the claim accordingly and, if possible, include an analysis using the actual segmentation labels to compute foreground/background information content (entropy/MI) without imposing the threshold. This does not affect the downstream evaluation but would strengthen the motivation.
minor comments (5)
  1. [Abstract/Introduction] Typographical errors: "serveral", "suprisingly", "comparision", "Breifly", "profound". Please proofread the manuscript.
  2. [§2.2, Eq. (2)-(3)] The notation H(X_p) is used both for the average intensity function and for the volume height h; consider renaming to avoid confusion.
  3. [§3.1] The paper mentions "no data augmentation was used during this stage" but downstream fine-tuning details (augmentation, epochs, learning rate schedule) are not specified; add for reproducibility.
  4. [Fig. 3a] The y-axis is "Values (log scale)" but bars are not annotated with exact values; add numbers or a table for readability.
  5. [Fig. 6a] References [2] and [18] are cited as "random" and "local" masking; ensure the exact implementations (e.g., masking ratio protocol) match and state the source of the numbers.

Circularity Check

0 steps flagged

No significant circularity: the HU-based masking strategy is an empirical method whose downstream gains are measured on independent benchmarks; the threshold and masking ratio are hyperparameters, not derivation inputs.

full rationale

The paper does not present a formal derivation whose output is equivalent to its input. The core proposal is an empirical masking strategy: subvolumes are selected by a Hounsfield-unit threshold (Eq. 2-3), a pretext model reconstructs the volume (Eq. 5), and downstream segmentation Dice is measured on five public datasets. The threshold λ=0.1 is a hyperparameter inferred from 40 CT images, not a quantity that is then 'predicted' by the model. The ROI analysis in Sec. 2.1 motivates the threshold by comparing entropy/complexity/MI of object versus background regions; even if this comparison is somewhat aligned with the threshold, it is not the same as the downstream evaluation and does not force the reported Dice scores. The masking ratio 0.6 is selected by a sweep on BTCV (Sec. 3.3, Fig. 6b), and the same split is used for the headline BTCV result; this is an experimental-design concern about potential selection bias, not a circularity in the derivation chain because the reported Dice is not mathematically determined by the masking ratio. The single self-citation (SDSL, ref. [11]) appears only as a baseline and is not load-bearing. The acknowledged limitation in Sec. 4 that a single global threshold may not be optimal is an honest caveat, not evidence of circularity. Overall, the central claim is self-contained against external benchmarks and no step reduces by construction to a fitted parameter or self-citation.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

The central contribution is a masking rule controlled by one fitted threshold and one masking ratio. No new entities are introduced. The method relies on several domain assumptions about HU distributions and transferability to MRI.

free parameters (3)
  • Foreground intensity threshold lambda = 0.1
    Chosen via analysis of 40 CT images (Sec 2.2) to separate anatomical structures from background; used globally across all CT and MRI datasets.
  • Masking ratio = 0.6
    Selected via ablation on BTCV (Fig6b); the best ratio on the evaluation dataset is then used to produce the reported BTCV result.
  • Subvolume size = 16x16x16
    Design choice based on prior subvolume partitioning works; not fitted but affects the granularity of foreground selection.
axioms (5)
  • domain assumption CT intensities are acquired in HU and normalized such that foreground organs mostly have values >= 0.1
    Used to define MF in Eq (2) and to justify foreground selection in Sec 1-2. If normalization differs across scanners, the threshold breaks.
  • domain assumption Background regions (air, fluid) carry no diagnostically relevant information for segmentation
    Core intuition in Sec 1; supported only by entropy/MI analysis on Flare22 (Fig3), not across all datasets.
  • domain assumption Mean subvolume HU is a sufficient statistic for foreground/background membership
    Eq (3) reduces each 16^3 subvolume to its mean; subvolumes with mixed tissue and air could be misclassified. No analysis of this proxy's accuracy.
  • domain assumption A threshold fitted on CT transfers to MRI (BraTS)
    BraTS results in Table 1 are produced with the same lambda=0.1 although MRI intensities are not Hounsfield units; no normalization or threshold adaptation is described.
  • domain assumption Entropy/complexity/MI analysis on Flare22 generalizes to other CT datasets
    The ROI analysis (Sec 2.1) is performed only on Flare22, then used to motivate the method for BTCV, MM-WHS, AMOS, and BraTS.

pith-pipeline@v1.3.0-alltime-deepseek · 6781 in / 16817 out tokens · 157032 ms · 2026-08-04T22:01:35.564475+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of HU-based Foreground Masking for 3D Medical Masked Image Modeling." pith.science (2026). https://pith.science/paper/TAHABPE6

@misc{pith2026250907534,
  author       = {Pith},
  title        = {Pith review of: HU-based Foreground Masking for 3D Medical Masked Image Modeling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TAHABPE6}},
  note         = {Machine review of arXiv:2509.07534}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

While Masked Image Modeling (MIM) has revolutionized fields of computer vision, its adoption in 3D medical image computing has been limited by the use of random masking, which overlooks the density of anatomical objects. To address this limitation, we enhance the pretext task with a simple yet effective masking strategy. Leveraging Hounsfield Unit (HU) measurements, we implement an HU-based Foreground Masking, which focuses on the intensity distribution of visceral organs and excludes non-tissue regions, such as air and fluid, that lack diagnostically meaningful features. Extensive experiments on five public 3D medical imaging datasets demonstrate that our masking consistently improves performance, both in quality of segmentation and Dice score (BTCV:~84.64\%, Flare22:~92.43\%, MM-WHS:~90.67\%, Amos22:~88.64\%, BraTS:~78.55\%). These results underscore the importance of domain-centric MIM and suggest a promising direction for representation learning in medical image segmentation. Implementation is available at github.com/AISeedHub/SubFore/.

Figures

Figures reproduced from arXiv: 2509.07534 by Anh Le, Gwang-Hyun Yu, Heonzoo Lee, Jin-Ho Jang, Jin Lee, Jin-Sul Kim, Jin-Young Kim, Kun-Yung Kim, Vu Dang, Zahid Rahman.

Figure 1
Figure 1. Figure 1: Brightness distribution. (a) An abdomen CT image with semantic seg [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: HU-based Foreground Masking framework. Given an input CT volume, [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: ROI analysis: (a) Comparison of mutual information, entropy, and com [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Dice Score (%) segmentation results across various datasets and back [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Segmentation quality comparison on Flare22 dataset. Our method shows [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Ablation Study. using a masking ratio of 0.6. Compared to random[2], and local[18] masking, our method yielded the significant improvement (0.75% in DSC), profound the importance of semantic-driven approaches in designing masking methods. 4 Conclusions Anatomical intensity in medical images offer valuable cues that can be leveraged for pretraining phase in Masked Image Modeling. In this study, we employ HU… 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

25 extracted references · 15 canonical work pages · 1 internal anchor

  1. [1]

    arXiv preprint arXiv:2106.08254 (2021)

    Bao, H., Dong, L., Piao, S., Wei, F.: Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254 (2021)

    Pith/arXiv arXiv 2021

  2. [2]

    In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision

    Chen, Z., Agarwal, D., Aggarwal, K., Safta, W., Balan, M.M., Brown, K.: Masked image modeling advances 3d medical image analysis. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1970– 1980 (2023)

    work page 1970

  3. [3]

    Journal of digital imaging 26, 1045–1057 (2013)

    Clark, K., Vendt, B., Smith, K., Freymann, J., Kirby, J., Koppel, P., Moore, S., Phillips, S., Maffitt, D., Pringle, M., et al.: The cancer imaging archive (tcia): main- taining and operating a public information repository. Journal of digital imaging 26, 1045–1057 (2013)

    work page 2013

  4. [4]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention

    Dai, P., Ou, Y., Yang, Y., Liu, D., Hashimoto, M., Jinzaki, M., Miyake, M., Suzuki, K.: Sasamim: Synthetic anatomical semantics-aware masked image modeling for colon tumor segmentation in non-contrast abdominal computed tomography. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 567–578. Springer (2024)

    work page 2024

  5. [5]

    In: International MICCAI brainlesion workshop

    Hatamizadeh, A., Nath, V., Tang, Y., Yang, D., Roth, H.R., Xu, D.: Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images. In: International MICCAI brainlesion workshop. pp. 272–284. Springer (2021)

    2021

  6. [6]

    In: Proceedings of the IEEE/CVF winter conference on applications of computer vi- sion

    Hatamizadeh, A., Tang, Y., Nath, V., Yang, D., Myronenko, A., Landman, B., Roth, H.R., Xu, D.: Unetr: Transformers for 3d medical image segmentation. In: Proceedings of the IEEE/CVF winter conference on applications of computer vi- sion. pp. 574–584 (2022)

    2022

  7. [7]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    He, K., Chen, X., Xie, S., Li, Y., Dollár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000–16009 (2022)

    2022

  8. [8]

    Advances in neural information processing systems 35, 36722–36732 (2022)

    Ji, Y., Bai, H., Ge, C., Yang, J., Zhu, Y., Zhang, R., Li, Z., Zhanng, L., Ma, W., Wan, X., et al.: Amos: A large-scale abdominal multi-organ benchmark for versatile medical image segmentation. Advances in neural information processing systems 35, 36722–36732 (2022)

    work page 2022

  9. [9]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Kong, X., Zhang, X.: Understanding masked image modeling via learning occlu- sion invariant feature. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6241–6251 (2023)

    2023

  10. [10]

    In: Proc

    Landman, B., Xu, Z., Igelsias, J., Styner, M., Langerak, T., Klein, A.: Miccai multi- atlas labeling beyond the cranial vault–workshop and challenge. In: Proc. MICCAI 10 J. Lee et al. Multi-Atlas Labeling Beyond Cranial Vault—Workshop Challenge. vol. 5, p. 12 (2015)

    work page 2015

  11. [11]

    IEEE Access (2024)

    Lee, J., Vu, D.T., Yu, G., Kim, J., Kim, K., Kim, J.: Sdsl: Spectral distance scaling loss pretraining swinunetr for 3d medical image segmentation. IEEE Access (2024)

    work page 2024

  12. [12]

    Advances in Neural Information Pro- cessing Systems35, 14290–14302 (2022)

    Li, G., Zheng, H., Liu, D., Wang, C., Su, B., Zheng, C.: Semmae: Semantic-guided masking for learning masked autoencoders. Advances in Neural Information Pro- cessing Systems35, 14290–14302 (2022)

    work page 2022

  13. [13]

    The Lancet Digital Health6(11), e815–e826 (2024)

    Ma, J., Zhang, Y., Gu, S., Ge, C., Ma, S., Young, A., Zhu, C., Yang, X., Meng, K., Huang, Z., Zhang, F., Pan, Y., Huang, S., Wang, J., Sun, M., Zhang, R., Jia, D., Choi, J.W., Alves, N., de Wilde, B., Koehler, G., Lai, H., Wang, E., Wiesen- farth, M., Zhu, Q., Dong, G., He, J., He, J., Yang, H., Huang, B., Lyu, M., Ma, Y., Guo, H., Xu, W., Maier-Hein, K.,...

    work page 2024

  14. [14]

    In: 2023 IEEE 33rd International Workshop on Machine Learning for Signal Pro- cessing (MLSP)

    Østmo, E.A., Wickstrøm, K.K., Radiya, K., Kampffmeyer, M.C., Jenssen, R.: View it like a radiologist: Shifted windows for deep learning augmentation of ct images. In: 2023 IEEE 33rd International Workshop on Machine Learning for Signal Pro- cessing (MLSP). pp. 1–6. IEEE (2023)

    work page 2023

  15. [15]

    Medical image analysis 42, 1–13 (2017)

    Setio, A.A.A., Traverso, A., De Bel, T., Berens, M.S., Van Den Bogaard, C., Cerello, P., Chen, H., Dou, Q., Fantacci, M.E., Geurts, B., et al.: Validation, com- parison, and combination of algorithms for automatic detection of pulmonary nod- ules in computed tomography images: the luna16 challenge. Medical image analysis 42, 1–13 (2017)

    2017

  16. [16]

    arXiv preprint arXiv:1902.09063 (2019)

    Simpson, A.L., Antonelli, M., Bakas, S., Bilello, M., Farahani, K., Van Ginneken, B., Kopp-Schneider, A., Landman, B.A., Litjens, G., Menze, B., et al.: A large an- notated medical image dataset for the development and evaluation of segmentation algorithms. arXiv preprint arXiv:1902.09063 (2019)

    Pith/arXiv arXiv 1902

  17. [17]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Tang, Y., Yang, D., Li, W., Roth, H.R., Landman, B., Xu, D., Nath, V., Hatamizadeh, A.: Self-supervised pre-training of swin transformers for 3d med- ical image analysis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 20730–20740 (2022)

    2022

  18. [18]

    Disruptive Autoencoders: Leveraging Low-level features for 3D Medical Image Pre-training

    Valanarasu, J.M.J., Tang, Y., Yang, D., Xu, Z., Zhao, C., Li, W., Patel, V.M., Landman, B., Xu, D., He, Y., et al.: Disruptive autoencoders: Leveraging low-level features for 3d medical image pre-training. arXiv preprint arXiv:2307.16896 (2023)

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  19. [19]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Wald, T., Ulrich, C., Lukyanenko, S., Goncharov, A., Paderno, A., Miller, M., Maerkisch, L., Jaeger, P., Maier-Hein, K.: Revisiting mae pre-training for 3d med- ical image segmentation. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 5186–5196 (2025)

    work page 2025

  20. [20]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Wang, H., Song, K., Fan, J., Wang, Y., Xie, J., Zhang, Z.: Hard patches mining for masked image modeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10375–10385 (2023)

    work page 2023

  21. [21]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention

    Wang, Y., Li, Z., Mei, J., Wei, Z., Liu, L., Wang, C., Sang, S., Yuille, A.L., Xie, C., Zhou, Y.: Swinmm: masked multi-view with swin transformers for 3d medical image segmentation. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 486–496. Springer (2023)

    work page 2023

  22. [22]

    arXiv preprint arXiv:2410.09890 (2024)

    Wu, L., Zhuang, J., Chen, H.: Large-scale 3d medical image pre-training with geometric context priors. arXiv preprint arXiv:2410.09890 (2024)

    Pith/arXiv arXiv 2024

  23. [23]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Wu,L.,Zhuang,J.,Chen,H.:Voco:Asimple-yet-effectivevolumecontrastivelearn- ing framework for 3d medical image analysis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22873–22882 (2024) HU-based Foreground Masking for 3D Medical Masked Image Modeling 11

    work page 2024

  24. [24]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Xie, Z., Zhang, Z., Cao, Y., Lin, Y., Bao, J., Yao, Z., Dai, Q., Hu, H.: Simmim: A simple framework for masked image modeling. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9653–9663 (2022)

    2022

  25. [25]

    IEEE transactions on pattern analysis and machine intelli- gence41(12), 2933–2946 (2018)

    Zhuang, X.: Multivariate mixture model for myocardial segmentation combining multi-source images. IEEE transactions on pattern analysis and machine intelli- gence41(12), 2933–2946 (2018)

    work page 2018