Pith. sign in

REVIEW 5 major objections 5 minor 53 references

ReCoSeg++:Extended Residual-Guided Cross-Modal Diffusion for Brain Tumor Segmentation

T0 review · 5 major / 5 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 Synthesizing T1ce MRI and feeding the synthesis error back as an attention map lifts whole-tumor Dice to 93.02% and IoU to 86.7% on BraTS 2021.

desk verdict Internal contradictions in the T1ce-absence claim and the reported numbers make this incremental extension untrustworthy as written, but the underlying idea and ablation work deserve a revised resubmission. read the letter →

arxiv 2508.01058 v1 pith:2NL6FPZH submitted 2025-08-01 eess.IV cs.CV

Sara Yavari , Rahul Nitin Pandya , Jacob Furst This is my paper
classification eess.IVcs.CV
keywords braintumorsegmentationdiffusionmodelscross-modalsynthesisresidualmapssemi-supervisedlearningBraTS2021wholeU-Net
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper proposes ReCoSeg++, a two-stage, semi-supervised pipeline for whole-tumor segmentation in brain MRI. Stage 1 trains a conditional denoising diffusion probabilistic model (DDPM) to reconstruct the T1ce modality from FLAIR, T1, and T2 scans; Stage 2 builds a residual map as the absolute difference between the real and synthesized T1ce images, concatenates it with the three native modalities, and feeds the four-channel input to a lightweight 2D U-Net trained with a BCE-plus-Dice loss. The paper reports a Dice score of 93.02% and an IoU (intersection over union) of 86.7% on BraTS 2021, outperforming UNet2D, UNet3D, DDMM-Synth, and the predecessor ReCoSeg on the same or earlier BraTS data. The claimed significance is that residual-guided attention improves boundary localization and scales to a larger, more heterogeneous multi-center dataset while keeping the segmenter computationally light and reducing reliance on dense annotations.

What carries the argument

The central object is the residual map $R$ of Eq. 7, the pixel-wise absolute difference between real and DDPM-synthesized T1ce images; it functions as an error-aware attention map whose high values mark reconstruction uncertainty that the authors associate with tumor-bearing regions. The segmentation machinery is the four-channel input of Eq. 8, $X_{\mathrm{seg}} = \mathrm{Concat}(x_{\mathrm{FLAIR}}, x_{\mathrm{T1}}, x_{\mathrm{T2}}, R)$, fed to a lightweight 2D U-Net trained with a combined BCE and Dice loss. Supporting components are slice-level filtering that drops axial slices without tumor signal, min-max normalization of residuals, threshold calibration that selects $\tau = 0.3$ for binarizing sigmoid outputs, and a plateau-based learning-rate schedule for the diffusion model.

What would settle it

Run the trained pipeline on BraTS 2021 with the real T1ce channel removed and the residual set to zero; if Dice falls to the no-residual baseline, the claim that residuals drive the gain is refuted. Separately, compute Dice over all axial slices, including tumor-free ones, and check whether the 93.02% headline survives full-volume evaluation.

Watch

Extended reading notes

Core claim

The central claim is that the residual map $R = |\hat{x}^{\mathrm{T1ce}}_0 - x^{\mathrm{T1ce}}_0|$, computed between the real T1ce scan and the T1ce synthesized by a conditional DDPM, acts as a soft attention prior that localizes tumor tissue and makes a lightweight 2D U-Net competitive with heavier supervised segmenters. Formally, the segmentation input is $X_{\mathrm{seg}} = \mathrm{Concat}(x_{\mathrm{FLAIR}}, x_{\mathrm{T1}}, x_{\mathrm{T2}}, R)$, and the U-Net is trained with a hybrid BCE and Dice loss on whole-tumor masks. On BraTS 2021 the paper reports 93.02% Dice and 86.7% IoU, beating UNet2D (87.3%/81.0%), UNet3D (88.1%/81.7%), DDMM-Synth (90.9%/85.1%), and ReCoSeg (91.2%/83.6%). Ablations attribute roughly +1.2 Dice to dynamic diffusion residuals over static error maps and roughly +0.8 Dice to threshold calibration at $\tau = 0.3$.

Load-bearing premise

The load-bearing premise is that the real T1ce scan is available at inference time, since Eq. 7 defines the residual as the absolute difference between real and synthesized T1ce and Eq. 8 feeds that residual to the segmenter; a second premise is that evaluating only tumor-bearing slices keeps the 93.02% Dice comparable to full-volume baselines.

Editorial extensions

If this is right

  • If the central claim holds, whole-tumor segmentation on BraTS 2021 reaches 93.02% Dice and 86.7% IoU with a lightweight 2D U-Net, so heavy 3D context is not required when a synthesis-residual attention map is available.
  • On BraTS 2021 the residual-guided lightweight U-Net outperforms both fully supervised 2D and 3D U-Nets and the diffusion-synthesis baseline DDMM-Synth, implying that synthesis error carries information beyond the native MRI channels.
  • The reported ablation gains (+1.2 Dice for dynamic over static residuals, +0.8 Dice for threshold calibration) give concrete levers for reproducing or improving the pipeline.
  • The framework's modularity means the diffusion synthesizer and the segmenter can be upgraded independently, which matters for clinical deployment where compute is limited.
  • The method's scale-up from ReCoSeg on BraTS 2020 to ReCoSeg++ on the larger, more heterogeneous BraTS 2021 suggests the residual-guided recipe transfers across datasets.

Reading between the lines

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

  • Testable extension: replace the real T1ce in Eq. 7 with a second independently sampled synthetic T1ce, so the residual is the difference between two synthesized scans; if Dice holds up, the pipeline becomes genuinely T1ce-free, which the equations as written do not support.
  • Because slices without tumor signal are excluded before training and evaluation, the reported 93.02% Dice is conditional on informative slices; a full-volume evaluation that scores every axial slice would be a stricter and more comparable test against baselines.
  • The residual-as-attention mechanism is a general recipe: any contrast-enhanced sequence that marks pathology could be synthesized from non-enhanced sequences, and its synthesis error could seed a segmenter for other lesion types; the paper only demonstrates this for whole-tumor segmentation in brain MRI.
  • A head-to-head comparison with ReCoSeg on identical BraTS 2021 splits, with per-subject paired statistics, would clarify whether the gain over the predecessor is due to the dynamic residual, the larger training set, or threshold calibration.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 5 minor

Summary. The paper proposes ReCoSeg++, a two-stage framework for whole-tumor segmentation in multi-modal brain MRI. Stage 1 trains a conditional DDPM to synthesize the T1ce modality from FLAIR, T1, and T2; a residual map R is defined in Eq. (7) as the absolute difference between synthesized and real T1ce. Stage 2 concatenates R with the three source modalities and feeds the four-channel input to a lightweight 2D U-Net trained with BCE plus Dice loss. The authors report Dice 93.02% and IoU 86.7% on BraTS 2021 and claim superior performance, scalability, and robustness to missing T1ce at inference.

Significance. Residual-guided cross-modal synthesis is a plausible idea, and a carefully executed study of it could interest the medical imaging community. The paper explicitly compares against its own predecessor and DDMM-Synth under a shared backbone, which is useful. However, the manuscript has load-bearing internal contradictions: Eqs. (7)-(8) are incompatible with the missing-T1ce claim; Table 1 contradicts the text and abstract; and the slice-level filtering protocol departs from the standard full-volume BraTS evaluation. The claimed elimination of ground-truth masks is also contradicted by Eq. (9), which uses ground-truth masks in the segmentation loss. No code, trained models, or statistical significance tests are provided. As a result, the headline claims are not established.

major comments (5)
  1. [Eq. (7), Eq. (8), Abstract, Section 6] The claim that ReCoSeg++ achieves accurate segmentation 'even in the absence of the T1ce modality during inference' is contradicted by the method definition. Eq. (7) defines R = |x_hat_T1ce_0 - x_T1ce_0|, which requires the real T1ce scan at inference; Eq. (8) then constructs Xseg = Concat(xFLAIR, xT1, xT2, R). If T1ce is unavailable, R is undefined and the U-Net cannot receive its four-channel input. No alternative definition or substitution rule for the missing-modality case is provided. This is load-bearing because the missing-T1ce capability is advertised in the Abstract and Section 6 as a key advantage for clinical deployment.
  2. [Table 1, Sections 4.3 and 4.4] Table 1 is internally inconsistent with the text. The table reports ReCoSeg++ on BraTS 2020 as Dice 89.8 and IoU 86.4, lower than ReCoSeg's 91.7/85.3, while Section 4.3 states that 'ReCoSeg++ achieves the highest performance, with a Dice score of 0.917 and an IoU of 0.853'—numbers that exactly match the ReCoSeg row. Section 4.4 then claims a Dice of approximately 93.02% on BraTS 2021 as superiority over ReCoSeg, but the only BraTS 2021 ReCoSeg number in the table is 91.2, not the 91.7 used in the Abstract. The reported comparison therefore does not support the abstract's claim of 'outperforming the ReCoSeg baseline on BraTS 2020.'
  3. [Sections 4.1 and 4.2] The evaluation protocol deviates from the standard BraTS full-volume protocol: slices without any tumor signal in the T1ce ground truth are excluded and top/bottom slices are discarded, so Dice and IoU are computed only on a filtered subset of slices. Published baselines are typically evaluated on whole volumes, so the reported numbers are not directly comparable. In addition, Section 4.1 says the top 26 and bottom 80 axial slices are discarded leaving 78 informative slices, which is arithmetically inconsistent with the standard 155-slice BraTS volumes (49 slices remain); this must be clarified and full-volume metrics reported.
  4. [Abstract, Section 3.2, Eq. (9)] The claim in the Abstract that the method 'eliminates the need for ground-truth masks for the segmentation objective' is not supported by the method. Eq. (9) defines L_seg = lambda1 * L_BCE(y_hat, y) + lambda2 * L_Dice(y_hat, y), which uses ground-truth masks y during training; no unsupervised or mask-free segmentation objective is defined. If the intended claim is only that masks are not needed in Stage 1 or at inference, that must be stated explicitly and the architecture description corrected.
  5. [Sections 4.2 and 4.4] The segmentation threshold tau = 0.3 and the final Dice/IoU numbers are both selected on the validation set: Section 4.2 describes a validation sweep over tau in {0.3, 0.4, 0.5}, and Section 4.4 reports the resulting validation-set performance as the headline result. This creates a circular selection-and-reporting procedure and likely overstates generalization; a held-out test set or external validation is required.
minor comments (5)
  1. [Sections 4.1 and 4.4] The dataset size for BraTS 2021 is given as 'over 400 subjects' in Section 4.1 but 'over 1,250 subjects' in Section 4.4; these numbers must be reconciled.
  2. [Abstract] The abstract contains a formatting typo: 'cross-modalsynthesisbyreconstructing' should be 'cross-modal synthesis by reconstructing.'
  3. [Section 4.4] The text describes residual maps as 'computed as voxel-wise absolute differences between ground truth and predicted T1ce,' but Eq. (7) defines them as differences between real and predicted T1ce; the terminology should be made consistent.
  4. [References] Reference [2] is cited as U-Net but is a remote-sensing captioning paper, and reference [3] appears to be TransBTS mis-cited as U-Net; the bibliography needs a careful correction pass.
  5. [Section 4.3] The sentence 'achieve accurate segmentation even without access to ground-truth T1ce at inference time' is confusing because the residual map requires real T1ce, not ground-truth T1ce, and the missing-modality claim is already contradicted by Eqs. (7)-(8).

Circularity Check

1 steps flagged · score 5.0 of 10

The headline Dice/IoU is partially circular: the segmentation threshold is tuned on the validation set, and the same validation set's scores are then reported as the result; the claimed ability to work without T1ce during inference is additionally contradicted by the method's own equations.

  1. fitted input called prediction [Section 4.2 (Implementation Details) and Table 1 caption]
    "We also incorporate threshold calibration, empirically selecting τ = 0.3 to binarize the sigmoid outputs, based on both visual assessment and validation performance (e.g., Dice and IoU). The threshold τ was empirically determined through a validation sweep across τ ∈ {0.3, 0.4, 0.5}, with τ = 0.3 yielding the best Dice-IoU tradeoff. ... Table 1: Comparison with baselines on BraTS2020 and BraTS2021 validation sets."

    The reported headline values (Dice 93.02%, IoU 86.7%) are measured on the BraTS2021 validation set, which is the same set used to select τ and to trigger early stopping (Section 4.2 also states 'early stopping based on the validation Dice score'). The reported score is therefore the result of optimizing a hyperparameter and a model-selection criterion on the evaluation set, not an independent prediction. This is the fitted-input-called-prediction pattern: the evaluation metric is used to choose the threshold, and the same metric on the same set is then presented as the method's performance.

full rationale

The paper is an empirical comparison rather than a mathematical derivation, so the only genuinely circular component is the validation-set selection-and-report loop. The residual idea is inherited from the authors' prior ReCoSeg [17], and [17] and DDMCIE [51] are cited baselines, but the central comparison is based on experiments reported in this paper rather than on any invoked uniqueness theorem or unverified prior result, so self-citation is not load-bearing. Two additional problems are correctness/consistency issues rather than circularity: Eq. 7 defines R as the absolute difference between real and synthesized T1ce and Eq. 8 makes R a required segmentation input, yet Section 6 claims the method works 'even in the absence of the T1ce modality during inference'; and the abstract claims to eliminate ground-truth masks while Eq. 9 trains the segmenter with y as 'the ground truth segmentation mask' and Section 4.1 filters slices using tumor ground truth. These undermine the stated claims but do not constitute derivational circularity. The threshold-tuning-on-validation issue is real and affects the headline numbers, so the circularity score is 5 rather than 0-2.

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

The central empirical claim rests on standard DDPM background, on the unproven mapping from T1ce reconstruction error to tumor location, and on a slice-filtered evaluation protocol that defines the reported Dice. Free parameters include the validation-tuned threshold, unreported loss weights, and an unspecified 'optimized' noise schedule. No new entities are introduced.

free parameters (4)
  • Binarization threshold tau = 0.3
    Chosen by a validation sweep over {0.3, 0.4, 0.5} in Sections 4.2 and 5; the reported Dice and IoU come from the same validation setting.
  • Reconstruction and segmentation loss weights lambda_1, lambda_2
    Weights in Eqs. 6 and 9 are not reported; they are hand-chosen hyperparameters that affect both stages.
  • Diffusion noise variance schedule
    Described as 'optimized' in Section 3.1 but no schedule is specified; the synthesis quality and therefore the residual maps depend on it.
  • Slice retention range = 78 axial slices (top 26 and bottom 80 removed)
    Section 4.1; this preprocessing choice defines the evaluation population and can inflate Dice by removing tumor-free slices.
assumptions (4)
  • standard math The DDPM forward and reverse processes, with the simplified noise-prediction objective, correctly model T1ce synthesis.
    Eqs. 1 through 4 follow Ho et al. [7]; accepted background used without modification.
  • domain assumption The residual between real and synthesized T1ce localizes tumor tissue.
    Sections 3.1 and 3.2 state that high reconstruction error 'likely contains tumor tissue'; no evidence is given that this holds across BraTS 2021 heterogeneity.
  • domain assumption Discarding slices without tumor signal preserves accuracy and generalization.
    Section 4.1 removes non-informative slices; the paper assumes this does not bias the reported Dice or the comparison to baselines.
  • domain assumption 2D slice-wise processing with 78 retained axial slices is a valid proxy for volumetric whole-tumor segmentation.
    Section 4.1 and 4.2 resample to 120x120 and drop slices; the paper compares against 3D baselines without clarifying whether all methods use the same filtered slice population.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ReCoSeg++:Extended Residual-Guided Cross-Modal Diffusion for Brain Tumor Segmentation." pith.science (2026). https://pith.science/paper/2NL6FPZH

@misc{pith2026250801058,
  author       = {Pith},
  title        = {Pith review of: ReCoSeg++:Extended Residual-Guided Cross-Modal Diffusion for Brain Tumor Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2NL6FPZH}},
  note         = {Machine review of arXiv:2508.01058}
}
abstract

Accurate segmentation of brain tumors in MRI scans is critical for clinical diagnosis and treatment planning. We propose a semi-supervised, two-stage framework that extends the ReCoSeg approach to the larger and more heterogeneous BraTS 2021 dataset, while eliminating the need for ground-truth masks for the segmentation objective. In the first stage, a residual-guided denoising diffusion probabilistic model (DDPM) performs cross-modal synthesis by reconstructing the T1ce modality from FLAIR, T1, and T2 scans. The residual maps, capturing differences between predicted and actual T1ce images, serve as spatial priors to enhance downstream segmentation. In the second stage, a lightweight U-Net takes as input the concatenation of residual maps, computed as the difference between real T1ce and synthesized T1ce, with T1, T2, and FLAIR modalities to improve whole tumor segmentation. To address the increased scale and variability of BraTS 2021, we apply slice-level filtering to exclude non-informative samples and optimize thresholding strategies to balance precision and recall. Our method achieves a Dice score of $93.02\%$ and an IoU of $86.7\%$ for whole tumor segmentation on the BraTS 2021 dataset, outperforming the ReCoSeg baseline on BraTS 2020 (Dice: $91.7\%$, IoU: $85.3\%$), and demonstrating improved accuracy and scalability for real-world, multi-center MRI datasets.

Figures

Figures reproduced from arXiv: 2508.01058 by the authors.

Figure 1
Figure 1. Overview of the proposed ReCoSeg framework for brain tumor [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Qualitative results of ReCoSeg++ for T1ce reconstruction: from left to [PITH_FULL_IMAGE:figures/full_fig_p012_2.png] view at source ↗
Figure 3
Figure 3. Qualitative segmentation results of ReCoSeg++ on the BraTS2020 [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Qualitative segmentation results of ReCoSeg++ on the BraTS2021 [PITH_FULL_IMAGE:figures/full_fig_p014_4.png]
Figure 5
Figure 5. Figure 5: Visualization of segmentation predictions under different binarization [PITH_FULL_IMAGE:figures/full_fig_p015_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

53 extracted references · 46 canonical work pages

  1. [4]

    Remote-sensing image captioning based on multilayer aggregated transformer.IEEE Geoscience and Remote Sensing Letters, 19:1–5, 2022

    Chenyang Liu, Rui Zhao, and Zhenwei Shi. Remote-sensing image captioning based on multilayer aggregated transformer.IEEE Geoscience and Remote Sensing Letters, 19:1–5, 2022

    work page 2022

  2. [24]

    Transbts: Multimodal brain tumor segmentation using transformer

    Wenqi Wang, Chen Chen, Mingyu Ding, Jingyun Li, Hongming Yu, and Shi Zha. Transbts: Multimodal brain tumor segmentation using transformer. In Proceedings of the International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI) , pages 109–119. Springer, 2021

    work page 2021

  3. [1]

    Nolte, and Mauricio Reyes

    Stefan Bauer, Roland Wiest, Lutz-P. Nolte, and Mauricio Reyes. A survey of mri-based medical image analysis for brain tumor studies.Physics in Medicine and Biology, 58(13):R97–R129, 2013

    work page 2013

  4. [5]

    Transbts: Multimodal brain tumor segmentation using transformer

    Wenxuan Wang, Chen Chen, Meng Ding, Hong Yu, Sen Zha, and Jiangyun Li. Transbts: Multimodal brain tumor segmentation using transformer. In International Conference on Medical Image Computing and Computer- Assisted Intervention (MICCAI) , pages 109–119. Springer, 2021

    work page 2021

  5. [6]

    DDMM-Synth: A Denoising Diffusion Model for Cross-modal Medical Image Synthesis with Sparse-view Measurement Embedding

    Yixiao Wang, Wenjia Zhang, Hanyu Li, Yule Qin, Ji Wu, and Yongchao Xu. Ddmm-synth: A denoising diffusion model for cross-modal medical image synthesis. arXiv preprint arXiv:2303.15770 , 2023

    work page Pith review arXiv 2023

  6. [7]

    J. Ho, A. Jain, and P. Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems , volume 33, pages 6840–6851. Curran Associates, Inc., 2020. 16

    work page 2020

  7. [8]

    Kawar et al

    B. Kawar et al. Imagic: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6007–6017. IEEE, 2023

    work page 2023

  8. [9]

    Gao et al

    S. Gao et al. Implicit diffusion models for continuous super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10021–10030. IEEE, 2023

    work page 2023

Show all 53 references
  1. [10]

    M. Sun, W. Huang, and Y. Zheng. Instance-aware diffusion model for gland segmentation in colon histology images. InProceedings of the Inter- national Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI), pages 662–672. Springer, 2023

    2023

  2. [11]

    Hierarchical text-conditional image generation with clip latents

    Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 , 2022

    2022 arXiv

  3. [12]

    Sara Mahdavi, Rapha Gontijo Lopes, et al

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to- image diffusion models with deep language understanding.arXiv preprint arXi...

    2022 arXiv

  4. [13]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684–10695. IEEE, 2022

    2022

  5. [14]

    Text-to-remote-sensing-image generation with structured generative adversarial networks.IEEE Geoscience and Remote Sensing Letters, 19:1–5, 2021

    Rui Zhao and Zhenwei Shi. Text-to-remote-sensing-image generation with structured generative adversarial networks.IEEE Geoscience and Remote Sensing Letters, 19:1–5, 2021

    2021

  6. [15]

    Generative adversarial networks

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde- Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM , 63(11):139–144, 2020

    2020

  7. [16]

    A two-stage cascade model with variational autoencoders and attention gates for mri brain tumor segmentation.Brain- lesion, pages 435–447, 2020

    Chenggang Lyu and Hai Shu. A two-stage cascade model with variational autoencoders and attention gates for mri brain tumor segmentation.Brain- lesion, pages 435–447, 2020

    2020

  8. [17]

    Recoseg: Residual guided cross-modal diffusion for efficient brain tumor segmentation

    Sara Yavari, Rahul Pandya, and Jacob Furst. Recoseg: Residual guided cross-modal diffusion for efficient brain tumor segmentation. InProceedings of the Medical Imaging with Deep Learning (MIDL) . DePaul University, Chicago, IL, USA, MIDL, 2025. Accepted for publication

    2025

  9. [18]

    Corrdiff: Corrective diffusion model for accurate mri brain tumor segmentation.IEEE Journal of Biomedical and Health Informatics , 28(3):1587, 2024

    Wenqing Li, Wenhui Huang, and Yuanjie Zheng. Corrdiff: Corrective diffusion model for accurate mri brain tumor segmentation.IEEE Journal of Biomedical and Health Informatics , 28(3):1587, 2024. 17

    2024

  10. [19]

    Chowdhury

    Aparna Pai, Rohil Shetty, Brendan Hodis, and Yuvraj S. Chowdhury. Magnetic resonance imaging physics, 2023. StatPearls [Internet]. Treasure Island (FL): StatPearls Publishing; Last updated: April 2, 2023

    2023

  11. [20]

    Automated quantitative tumour response assessment of mri in neuro-oncology with artificial neural networks: A multicentre, retrospective study.The Lancet Oncology, 2019

    Philipp Kickingereder, Fabian Isensee, Irada Tursunova, et al. Automated quantitative tumour response assessment of mri in neuro-oncology with artificial neural networks: A multicentre, retrospective study.The Lancet Oncology, 2019

    2019

  12. [21]

    Patterson, and Huixiao Hong

    Md Kamrul Hasan Khan, Wenjing Guo, Jie Liu, Fan Dong, Zoe Li, Tucker A. Patterson, and Huixiao Hong. Machine learning and deep learning for brain tumor mri image segmentation.Experimental Biology and Medicine (Maywood), 248(21):1974–1992, 2023

    1974

  13. [22]

    Chandramma, T

    Asma Alshuhail, Arastu Thakur, R. Chandramma, T. R. Mahesh, Ahlam Almusharraf, V. Vinoth Kumar, and Surbhi Bhatia Khan. Refining neural network algorithms for accurate brain tumor classification in mri imagery. BMC Medical Imaging, 24:118, 2024

    2024

  14. [23]

    Instance-aware diffusion model for gland segmentation in colon histology images

    Ming Sun, Weilin Huang, and Yefeng Zheng. Instance-aware diffusion model for gland segmentation in colon histology images. InProceedings of the International Conference on Medical Image Computing and Computer- Assisted Intervention (MICCAI) , pages 662–672. Springer, 2023

    2023

  15. [25]

    Canet: Context aware network for brain glioma seg- mentation

    Zongwei Liu, Guotai Wang, Lequan Yu, Yizhe Zhang, Xiaomeng Li, and Pheng-Ann Heng. Canet: Context aware network for brain glioma seg- mentation. IEEE Transactions on Medical Imaging , 40(7):1763–1777, Jul 2021

    2021

  16. [26]

    Unet++: A nested u-net architecture for medical image segmentation

    Zongwei Zhou, Md Mahfuzur Rahman Siddiquee, Nima Tajbakhsh, and Jianming Liang. Unet++: A nested u-net architecture for medical image segmentation. CoRR, abs/1807.10165, 2018

    2018 arXiv

  17. [27]

    Lienkamp, Thomas Brox, and Olaf Ronneberger

    Özgün Çiçek, Ahmed Abdulkadir, Soeren S. Lienkamp, Thomas Brox, and Olaf Ronneberger. 3d u-net: Learning dense volumetric segmentation from sparse annotation. InInternational Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI) , pages 424–432. Spr...

    2016

  18. [28]

    V-net: Fully convolutional neural networks for volumetric medical image segmentation

    Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In Proceedings of the 2016 Fourth International Conference on 3D Vision (3DV), pages 565–571. IEEE, 2016. 18

    2016

  19. [29]

    Hammerla, Bernhard Kainz, Ben Glocker, and Daniel Rueckert

    Ozan Oktay, Jo Schlemper, Loic Le Folgoc, Matthew Lee, Mattias Heinrich, Kazunari Misawa, Kensaku Mori, Steven McDonagh, Nils Y. Hammerla, Bernhard Kainz, Ben Glocker, and Daniel Rueckert. Attention u-net: Learning where to look for the pancreas.CoRR, abs/1804.03999, 2018

    2018 arXiv

  20. [30]

    Yuille, and Yuyin Zhou

    Jieneng Chen, Yongyi Lu, Qihang Yu, Xiangde Luo, Ehsan Adeli, Yan Wang, Le Lu, Alan L. Yuille, and Yuyin Zhou. Transunet: Transformers make strong encoders for medical image segmentation.CoRR, abs/2102.04306, 2021

    2021 arXiv

  21. [31]

    Swin-unet: Unet-like pure trans- former for medical image segmentation.CoRR, abs/2105.05537, 2022

    Huwaisong Cao, Yueyue Wang, Joy Iong Zong, Jianfei Yang, Ying Chen, Zhen Li, Maoqing Tian, and Li Zhang. Swin-unet: Unet-like pure trans- former for medical image segmentation.CoRR, abs/2105.05537, 2022

    2022 arXiv

  22. [32]

    Jaeger, Simon Kohl, Jakob Wasserthal, Gregor Köhler, Tobias Norajitra, Sebastian Wirkert, and Klaus H

    Fabian Isensee, Jens Petersen, Andre Klein, David Zimmerer, Paul F. Jaeger, Simon Kohl, Jakob Wasserthal, Gregor Köhler, Tobias Norajitra, Sebastian Wirkert, and Klaus H. Maier-Hein. nnu-net: A self-configuring method for deep learning-based biomedical image segmentation.Natur...

    2021

  23. [33]

    Extending nnu-net for brain tumor segmentation

    Huan Minh Luu and Sung-Hong Park. Extending nnu-net for brain tumor segmentation. In International MICCAI Brainlesion Workshop , volume 12963 of Lecture Notes in Computer Science , pages 173–186. Springer, 2021

    2021

  24. [34]

    Optimized u-net for brain tumor segmentation

    Michał Futrega, Alexandre Milesi, Michał Marcinkiewicz, and Pablo Ribalta. Optimized u-net for brain tumor segmentation. InInternational MICCAI Brainlesion Workshop, volume 12963 ofLecture Notes in Computer Science , pages 15–29. Springer, 2021

    2021

  25. [35]

    Redundancy reduction in semantic segmentation of 3d brain tumor mris.arXiv preprint, arXiv:2111.00742, 2021

    Md Mahfuzur Rahman Siddiquee and Andriy Myronenko. Redundancy reduction in semantic segmentation of 3d brain tumor mris.arXiv preprint, arXiv:2111.00742, 2021

    2021 arXiv

  26. [36]

    Generative medical segmentation, 2024

    Jiayu Huo, Xi Ouyang, Sébastien Ourselin, and Rachel Sparks. Generative medical segmentation, 2024. arXiv preprint arXiv:2403.17876

    2024 arXiv

  27. [37]

    Generative ai enables medical image seg- mentation in ultra low-data regimes, 2024

    Li Zhang, Basu Jindal, Ahmed Alaa, Robert Weinreb, David Wilson, Eran Segal, James Zou, and Pengtao Xie. Generative ai enables medical image seg- mentation in ultra low-data regimes, 2024. arXiv preprint arXiv:2408.17421

    2024 arXiv

  28. [38]

    Aswani and D

    K. Aswani and D. Menaka. A dual autoencoder and singular value decom- position based feature optimization for the segmentation of brain tumor from mri images. BMC Medical Imaging, 21:82, 2021

    2021

  29. [39]

    Drm-vae: A dual residual multi variational auto-encoder for brain tumor segmentation with missing modalities

    Yian Zhu, Shaoyu Wang, Yun Hu, and Xiao Ma. Drm-vae: A dual residual multi variational auto-encoder for brain tumor segmentation with missing modalities. In 2021 IEEE 4th International Conference on Electronics and Communication Engineering (ICECE). IEEE, 2021. 19

    2021

  30. [40]

    Generative adversarial networks

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde- Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. arXiv preprint, 2014

    2014

  31. [41]

    H. C. Shin, N. A. Tenenholtz, J. K. Rogers, C. G. Schwarz, M. L. Senjem, J. L. Gunter, K. P. Andriole, and M. Michalski. Medical image synthesis for data augmentation and anonymization using generative adversarial networks. In Lecture Notes in Computer Science , volume 11037, ...

    2018

  32. [42]

    Generative adversarial network in medical imaging: A review.Medical Image Analysis, 2019

    Xin Yi, Ekta Walia, and Paul Babyn. Generative adversarial network in medical imaging: A review.Medical Image Analysis, 2019

    2019

  33. [43]

    Vox2vox: 3d-gan for brain tumour segmentation.arXiv preprint, 2020

    Marco Domenico Cirillo, David Abramian, and Anders Eklund. Vox2vox: 3d-gan for brain tumour segmentation.arXiv preprint, 2020

    2020

  34. [44]

    Promptable counterfac- tual diffusion model for unified brain tumor segmentation and generation with mris

    Yiqing Shen, Guannan He, and Mathias Unberath. Promptable counterfac- tual diffusion model for unified brain tumor segmentation and generation with mris. arXiv preprint, 2024

    2024

  35. [45]

    Pohl, and Yu Zhang

    Haokai Zhao, Haowei Lou, Lina Yao, Wei Peng, Ehsan Adeli, Kilian M. Pohl, and Yu Zhang. Diffusion models for computational neuroimaging: A survey. arXiv preprint, 2024

    2024

  36. [46]

    Jonas Wolleb, Robin Sandkühler, Florian Bieder, Pietro Valmaggia, and Pascal C. Cattin. Diffusion models for implicit image segmentation ensem- bles. In International Conference on Medical Imaging with Deep Learning (MIDL). PMLR, 2022

    2022

  37. [47]

    Accelerating diffusion models via pre-segmentation diffusion sampling for medical image segmentation.arXiv preprint, 2022

    Xutao Guo, Yanwu Yang, Chenfei Ye, Shang Lu, Yang Xiang, and Ting Ma. Accelerating diffusion models via pre-segmentation diffusion sampling for medical image segmentation.arXiv preprint, 2022

    2022

  38. [48]

    Med- segdiff: Medical image segmentation with diffusion probabilistic model

    Jiachen Wu, Hao Fang, Yuying Zhang, Yubo Yang, and Yufeng Xu. Med- segdiff: Medical image segmentation with diffusion probabilistic model. arXiv preprint, 2022

    2022

  39. [49]

    Medsegdiff-v2: Diffusion based medical image segmentation with transformer

    Junde Wu, Wei Ji, Rao Fu, Huazhu Fu, Min Xu, Yueming Jin, and Yanwu Xu. Medsegdiff-v2: Diffusion based medical image segmentation with transformer. arXiv preprint, 2023

    2023

  40. [50]

    Segdiff: Image segmentation with diffusion probabilistic models

    Tomer Amit, Eliya Nachmani, Tom Shaharbany, and Lior Wolf. Segdiff: Image segmentation with diffusion probabilistic models. arXiv preprint, 2021

    2021

  41. [51]

    Sara Yavari, Rahul Nitin Pandya, and Jacob Furst. Dmcie: Diffusion model with concatenation of inputs and errors to improve the accuracy of the seg- mentation of brain tumors in mri images.arXiv preprint arXiv:2507.00983 , Jul 2025. 20

    2025 arXiv

  42. [52]

    Lightm-unet: Mamba assists in lightweight unet for medical image segmentation

    Weibin Liao, Yinghao Zhu, Xinyuan Wang, Chengwei Pan, Yasha Wang, and Liantao Ma. Lightm-unet: Mamba assists in lightweight unet for medical image segmentation. arXiv preprint arXiv:2403.05246 , 2024

    2024 arXiv

  43. [53]

    Zou and et al

    Kelly H. Zou and et al. Statistical validation of image segmentation quality based on a spatial overlap index.Academic Radiology, 11(2):178–189, 2004

    2004

  44. [54]

    Optimization for medical image segmentation: Theory and practice when evaluating with dice score or jaccard index.IEEE Transactions on Medical Imaging , 39(11):3679–3690, 2020

    Tom Eelbode and et al. Optimization for medical image segmentation: Theory and practice when evaluating with dice score or jaccard index.IEEE Transactions on Medical Imaging , 39(11):3679–3690, 2020

    2020

  45. [55]

    U-net: Convolutional networks for biomedical image segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. InInternational Conference on Medical Image Computing and Computer-Assisted Intervention , pages 234–241. Springer, 2015. 21

    2015

Pith tools

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