Pith. sign in

REVIEW 4 major objections 6 minor 31 references

UniSegDiff: Boosting Unified Lesion Segmentation via a Staged Diffusion Model

T0 review · 4 major / 6 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 Staged diffusion reaches 84.4 mDice across six lesion datasets

desk verdict Staged diffusion training idea is worth a second look, but the empty-mask handling in Lung Infection and inconsistent aggregates put the SOTA claim on hold. read the letter →

arxiv 2507.18362 v1 pith:ADR2YLHH submitted 2025-07-24 eess.IV cs.CV

Yilong Hu , Shijie Chang , Lihe Zhang , Feng Tian , Weibing Sun , Huchuan Lu This is my paper
classification eess.IVcs.CV
keywords unifiedlesionsegmentationdiffusionprobabilisticmodelstagedtrainingandinferencemedicalimagemulti-modalimagingdualcross-attentionSTAPLEensemblefusiondenoising
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

UniSegDiff is a diffusion-based framework for segmenting lesions across different organs and imaging modalities in one unified model. The paper argues that standard diffusion training spreads a model's attention unevenly across noise timesteps, slowing convergence and hurting accuracy, and it responds by splitting training and inference into three stages with different prediction targets so the network stays attentive at every timestep. A feature-extraction subnetwork pre-trained for segmentation is frozen during denoising, narrowing the gap between modalities, and multiple sampled masks are fused with STAPLE. On six public datasets the method reports consistent improvements over prior single-task and unified baselines, with a unified mDice of 84.4 and roughly 25 hours of training.

What carries the argument

The machinery is a three-stage training schedule with dynamic prediction targets and loss weights. In the Rapid Segmentation stage all timesteps are collapsed to $t=999$ and the network predicts the original mask $x_0$ with loss weights $\alpha:\beta=1:3$; in the Probabilistic Modeling stage ($299<t\le599$) it predicts both $x_0$ and noise $\epsilon$ with equal weight; in the Denoising Refinement stage all timesteps collapse to $t=0$ and it predicts $\epsilon$ with weights $3:1$. At inference the network jumps from pure noise at $t=999$ to $X_{599}$ in one step, uses DDIM steps of size 30, and then single-steps each branch from $X_{299}$ to $X_0$, producing 20 masks that are merged by the STAPLE algorithm. A second UNet, called CFENet, is pre-trained for segmentation on the unified dataset and frozen, with conditional features injected through dual cross-attention blocks in which conditional and denoising features alternate as queries.

What would settle it

Train UniSegDiff with the Rapid Segmentation stage sampling timesteps uniformly in the range $(599,999]$ instead of fixing all of them at $t=999$, keeping everything else identical. If unified mDice stays at or above 84.4 with the same 25-hour budget, the claim that collapsing to the boundary maintains high attention is not doing the work, while a sharp drop would confirm that the boundary-only schedule is load-bearing. A second check is to run staged inference starting from $t=800$ or $t=400$ and compare segmentation quality against the standard $t=999$ start.

Watch

Extended reading notes

Core claim

The central claim is that a staged diffusion schedule, predicting the clean mask in the high-noise Rapid Segmentation stage, predicting both noise and mask in the middle Probabilistic Modeling stage, and predicting noise in the low-noise Denoising Refinement stage, keeps the denoising network's gradient attention high across all timesteps and overcomes the slow convergence and suboptimal solutions seen when a diffusion model always predicts either noise or the mask. On top of this, the paper claims that pre-training the conditional feature extractor as a segmentation network and freezing it during denoising substantially reduces cross-modality feature confusion, enabling one model to segment wet AMD in OCT, brain tumors in MRI, adenocarcinoma in pathology, colon polyps in endoscopy, lung infections in CT, and breast lesions in ultrasound with no noticeable drop relative to individual-task training. With staged inference and STAPLE fusion of 20 sampled masks, the framework reports state-of-the-art mDice on all six individual tasks and on the unified task.

Load-bearing premise

The staged schedule trains the network only at the boundary timesteps $t=999$ and $t=0$ for the two outer stages, yet inference runs DDIM through intermediate timesteps the network never sees in those stages, so the gains rely on the unstated assumption that denoising skill transfers from the boundaries to all intermediate timesteps.

Editorial extensions

If this is right

  • Training time drops to about 25 hours and inference to as few as 11 steps, roughly 10 times faster than DDIM sampling, while unified mDice reaches 84.4.
  • Staged training alone raises unified mDice from 77.0 to 80.5, and the full pipeline with pre-training, dual cross-attention, and fusion reaches 85.3 in the component ablation.
  • A single unified model handles six organs and imaging modalities without the large performance collapse that discriminative baselines show on the unified task.
  • Threshold selection matters: the ablation finds $t=600$ and $t=300$ optimal, and moving either boundary can lower unified mDice to 82.1 or 82.7.
  • The staged schedule is not just a speed trick; the paper reports that it also improves accuracy over uniform noise prediction, uniform mask prediction, and one-step denoising under the same training budget.

Reading between the lines

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

  • If the uneven-attention diagnosis is right, the same three-stage principle should transfer to other conditional diffusion tasks, such as image-to-image translation or panoptic segmentation, where predicting the clean target at high noise and the residual at low noise could shorten training.
  • Training only at the boundary timesteps $t=999$ and $t=0$ may effectively implement a two-point curriculum; a direct test would be whether the model generalizes to intermediate timesteps better than a model trained on a few uniformly sampled timesteps.
  • The authors note that roughly one-third of CT lung masks are empty, so cleaning that data would likely raise both the Lung Infection scores and the unified average, meaning the reported margin should be re-checked after data cleaning.
  • The 0.9-point fusion gain comes from combining 20 diverse samples; a cheaper extension would use fewer samples with a deterministic noise schedule to see how much of that gain is diversity rather than averaging.
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

4 major / 6 minor

Summary. The manuscript proposes UniSegDiff, a diffusion-based framework for unified lesion segmentation across six datasets spanning OCT, MRI, pathology, endoscopy, CT, and ultrasound. The method combines a pre-trained frozen conditional feature extraction network with a two-decoder denoising UNet, a staged training schedule that alternates the prediction target between the original mask and the noise, a staged DDIM-like inference procedure, and STAPLE-based fusion of 20 sampled masks. The authors report state-of-the-art results on both individual and unified lesion segmentation tasks, with a unified-task average mDice of about 84.4 and a training time of about 25 hours. The paper also provides ablations of the denoising strategy, the stage thresholds, the architectural components, and the time efficiency.

Significance. If the reported results are substantiated, UniSegDiff is a practically useful contribution: it addresses a real unification problem in medical lesion segmentation, releases code, evaluates on six public benchmarks, and claims a large reduction in training time relative to prior diffusion-based methods. The staged-training idea and the use of a frozen pre-trained segmentation network to align conditional and denoising features are both worth investigating further. However, the manuscript currently contains load-bearing gaps: the Lung Infection dataset is admitted to contain roughly one-third empty masks, the staged-training timestep description is internally inconsistent, and the aggregate tables do not match the per-task numbers. These issues prevent the empirical claims from being accepted as stated.

major comments (4)
  1. [Section 3.2, Table 3] The 'Defect analysis' paragraph admits that approximately one-third of the Lung Infection masks are empty and states that the authors 'will clean the data and re-validate the results in future work.' Despite this, Table 3 reports Lung Infection results and includes this task in the unified-task averages, and the abstract and conclusion claim consistent state-of-the-art performance across all tasks. The paper never states whether empty-mask scans were excluded from training and evaluation, how an empty ground-truth mask is scored (e.g., whether empty-vs-empty counts as Dice=1), or whether every comparator used the same inclusion protocol. These choices can materially change the reported Dice values. Please either exclude Lung Infection from the headline unified results, provide results on cleaned data with an explicit protocol, or report a sensitivity analysis showing that the comparison is robust to empty-mask handling.
  2. [Section 2.2, Eq. (1)] The staged-training description is internally inconsistent as written. The Rapid Segmentation Stage is defined as (599 < t) and then all time steps are set to the maximum value (t = 999), so during that stage the model only trains at t=999. The Denoising Refinement Stage is defined as (t <= 299) and then all time steps are set to the minimum value (t = 0). Taken literally, the model trains on clean masks at t=0, where the noise prediction loss Ln is degenerate, and it never sees intermediate timesteps during the refinement stage, even though inference runs DDIM from t=299 down to t=0. If t=0 is a typo for t=299, the text must be corrected. In either case, the paper does not provide evidence that training only at boundary timesteps transfers to the intermediate timesteps used at inference; this transfer is load-bearing for the reported speed and accuracy gains, so an analysis or ablation demonstrating it is needed.
  3. [Tables 3, 4, and 7] The aggregate unified-task numbers in Tables 4 and 7 (mDice 84.4, mIoU 76.3) do not match the simple average of the six per-task unified results in Table 3. Averaging the Table 3 unified mDice values gives 84.9, and averaging the mIoU values gives 77.15. Since Section 3.3 states that 'we only present the average values of the metrics across all datasets,' a simple average is the natural reading, so this discrepancy cannot be attributed to a different weighting without an explicit statement. Please reconcile the tables, report the exact averaging protocol used for each aggregate number, and ensure that every aggregate claim is consistent with the per-task results.
  4. [Section 3.3, Table 5] The threshold ablation does not uniquely support the claimed optimal threshold. Table 5 shows that (600, 300) and (600, 200) both give exactly 84.4 mDice and 76.3 mIoU, so t=300 is not a unique optimum; the rows without a low-threshold label also make the grid ambiguous. Furthermore, the stage thresholds and the alpha:beta loss-weight ratios in Eq. (1) appear to be selected on the target benchmark itself, so the staged-training advantage in Table 4 is partly a hyperparameter-search result rather than an independent prediction. Please report how the thresholds and weights were selected (e.g., on a validation fold), include error bars for each threshold configuration, and discuss the sensitivity of the central claim to these choices.
minor comments (6)
  1. [Section 2.1] The phrase '2.5 UNet networks' is imprecise; the architecture should be described in terms of the number of encoders, decoders, and parameter-sharing relationships.
  2. [Section 2.2] The interval notation is ambiguous: '599 < t' should be written as an integer range (t >= 600) or (t in [600, 999]), and 't <= 299' should be matched with the threshold values used in Table 5.
  3. [Section 3.2] The paragraph titled 'Defect analysis' is really an acknowledgment of a data limitation; it should be moved to a limitations discussion and its implications should be reflected in the abstract and conclusion claims.
  4. [Section 3.3, Table 5] The table rows do not consistently specify both the high and low thresholds; for example, the row '700' needs a low-threshold entry, and '600 300' and '600 200' produce identical metrics in the text-adjacent rows.
  5. [Section 3.3, Table 6] The column header 'Pre-Tra' and the row labels 'Staged Pre-Tra DCA Fusion' are abbreviations that should be expanded in the caption or table header for readability.
  6. [Throughout] The word 'significantly' is used in the abstract and Section 3.2 without any statistical significance testing; given the reported standard deviations, paired tests across the 4 cross-validation folds or confidence intervals would substantiate the claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: UniSegDiff's staged design is an empirical contribution validated by ablations and external baselines; admitted benchmark issues are validity concerns, not derivation loops.

full rationale

UniSegDiff's claimed derivation chain is an empirical method evaluation, not a formal derivation, and I find no step in which a predicted quantity is equivalent by construction to an input. The staged thresholds (600/300) and loss weights are selected via ablations (Table 5) on the unified benchmark; this is conventional hyperparameter search and is reported as such ('the optimal low threshold is t = 300'), not a fitted parameter renamed as a prediction of the same benchmark. The central speed and accuracy claims are checked against external discriminative and diffusion baselines (Table 3), and the fusion and sampling components (STAPLE [27], DDIM [24]) are external, machine-independent algorithms. The only self-citation is the AMD-SD dataset reference [10], which is a data contribution by a co-author; it is not load-bearing for any derivation. The paper itself flags a limitation: 'a large number of masks being empty (approximately one-third of the dataset)' in Lung Infection and promises to clean and re-validate. That is a benchmark-validity concern and could affect Table 3 comparisons if empty-mask scoring differs, but it is not a circularity of the kind where an output reduces to an input. Similarly, training only at t=999 and t=0 while sampling intermediate timesteps with DDIM is a potential train/inference mismatch, but it is an empirical generalization question, not a self-referential loop. I therefore find no significant circularity.

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

The central claims rest on standard diffusion processes, on an unproven feature-alignment assumption for CFENet, and on a qualitative attention/gradient hypothesis. The reported method introduces several hand-tuned hyperparameters (stage thresholds, loss weights, sample counts, STAPLE settings) that are fit to the same benchmarks used for evaluation, which increases the risk that the measured gains are partly tuning artifacts.

free parameters (4)
  • Stage thresholds (high, low) = 600, 300
    Selected via ablation on the unified task (Table 5) to maximize mDice; 700 and 400 give lower scores, showing test-time tuning.
  • Loss weight ratios alpha:beta = 1:3, 1:1, 3:1
    Chosen by hand per stage in Section 2.2; no ablation or independent justification is given.
  • Number of sample repeats per branch = 10 mask, 10 noise, 20 total
    Fixed in staged inference (Section 2.2); affects the speed-accuracy tradeoff and is not swept in the paper.
  • STAPLE initial parameters = alpha=0.9, beta=0.1, prior=0.5, 20 iterations
    Initial confidence values and update iterations are hand-set in Section 2.4; no sensitivity analysis is reported.
assumptions (4)
  • standard math DDPM/DDIM forward and reverse processes are valid generative models for binary masks
    Relied on throughout Sections 2.1 and 2.2; standard results from Ho et al. and Song et al.
  • domain assumption CFENet pre-trained as a standalone segmentation model produces conditional features whose distribution matches the mask distribution and transfers to DNet
    Asserted in Section 2.3 without distributional evidence; this is the core justification for unified segmentation.
  • ad hoc to paper Average gradient magnitude across timesteps measures 'attention' and rebalancing it via staged targets improves convergence and accuracy
    Introduced in Section 1 with a qualitative plot; no formal or quantitative link between gradient magnitude, attention, and final accuracy is shown.
  • domain assumption STAPLE consensus of 20 diffusion samples improves accuracy and robustness
    Standard algorithm from Warfield et al., but the benefit here is empirical; Table 6 shows only a +0.9 mDice gain for the fusion component.

how reviews work

0 comments
Cite this review

Pith. "Pith review of UniSegDiff: Boosting Unified Lesion Segmentation via a Staged Diffusion Model." pith.science (2026). https://pith.science/paper/ADR2YLHH

@misc{pith2026250718362,
  author       = {Pith},
  title        = {Pith review of: UniSegDiff: Boosting Unified Lesion Segmentation via a Staged Diffusion Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ADR2YLHH}},
  note         = {Machine review of arXiv:2507.18362}
}
read the original abstract

The Diffusion Probabilistic Model (DPM) has demonstrated remarkable performance across a variety of generative tasks. The inherent randomness in diffusion models helps address issues such as blurring at the edges of medical images and labels, positioning Diffusion Probabilistic Models (DPMs) as a promising approach for lesion segmentation. However, we find that the current training and inference strategies of diffusion models result in an uneven distribution of attention across different timesteps, leading to longer training times and suboptimal solutions. To this end, we propose UniSegDiff, a novel diffusion model framework designed to address lesion segmentation in a unified manner across multiple modalities and organs. This framework introduces a staged training and inference approach, dynamically adjusting the prediction targets at different stages, forcing the model to maintain high attention across all timesteps, and achieves unified lesion segmentation through pre-training the feature extraction network for segmentation. We evaluate performance on six different organs across various imaging modalities. Comprehensive experimental results demonstrate that UniSegDiff significantly outperforms previous state-of-the-art (SOTA) approaches. The code is available at https://github.com/HUYILONG-Z/UniSegDiff.

Figures

Figures reproduced from arXiv: 2507.18362 by the authors.

Figure 1
Figure 1. The top-left corner shows the relationship between the model’s average [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

31 extracted references · 22 canonical work pages

  1. [1]

    https://medicalsegmentation.com/covid19/ (2020), accessed: June 25, 2025

    Covid-19 ct lung and infection segmentation dataset. https://medicalsegmentation.com/covid19/ (2020), accessed: June 25, 2025

    work page 2020

  2. [2]

    Data in brief28, 104863 (2020)

    Al-Dhabyani, W., Gomaa, M., Khaled, H., Fahmy, A.: Dataset of breast ultrasound images. Data in brief28, 104863 (2020)

    work page 2020

  3. [3]

    arXiv preprint arXiv:2112.00390 (2021)

    Amit, T., Shaharbany, T., Nachmani, E., Wolf, L.: Segdiff: Image segmentation with diffusion probabilistic models. arXiv preprint arXiv:2112.00390 (2021)

    arXiv 2021

  4. [4]

    arXiv preprint arXiv:2102.04306 (2021)

    Chen, J., Lu, Y., Yu, Q., Luo, X., Adeli, E., Wang, Y., Lu, L., Yuille, A.L., Zhou, Y.: Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306 (2021)

    arXiv 2021

  5. [5]

    PloS one10(10), e0140381 (2015)

    Cheng, J., Huang, W., Cao, S., Yang, R., Yang, W., Yun, Z., Wang, Z., Feng, Q.: Enhanced performance of brain tumor classification via tumor region augmentation and partition. PloS one10(10), e0140381 (2015)

    work page 2015

  6. [6]

    PloS one11(6), e0157112 (2016)

    Cheng, J., Yang, W., Huang, M., Huang, W., Jiang, J., Zhou, Y., Yang, R., Zhao, J., Feng, Y., Feng, Q., et al.: Retrieval of brain tumors by adaptive spatial pooling and fisher vector representation. PloS one11(6), e0157112 (2016)

    work page 2016

  7. [7]

    In: MICCAI

    Fan, D.P., Ji, G.P., Zhou, T., Chen, G., Fu, H., Shen, J., Shao, L.: Pranet: Par- allel reverse attention network for polyp segmentation. In: MICCAI. pp. 263–273. Springer (2020)

    2020

  8. [8]

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

    Hejrati, B., Banerjee, S., Glide-Hurst, C., Dong, M.: Conditional diffusion model with spatial attention and latent embedding for medical image segmentation. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 202–212. Springer (2024)

    work page 2024

Show all 31 references
  1. [9]

    Advances in neural information processing systems33, 6840–6851 (2020) 10 Authors Suppressed Due to Excessive Length

    Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems33, 6840–6851 (2020) 10 Authors Suppressed Due to Excessive Length

    2020

  2. [10]

    Scientific Data11(1), 1014 (2024)

    Hu, Y., Gao, Y., Gao, W., Luo, W., Yang, Z., Xiong, F., Chen, Z., Lin, Y., Xia, X., Yin, X., et al.: Amd-sd: An optical coherence tomography image dataset for wet amd lesions segmentation. Scientific Data11(1), 1014 (2024)

    2024

  3. [11]

    In: MultiMedia modeling: 26th international conference, MMM 2020, Daejeon, South Korea, January 5–8, 2020, proceedings, part II 26

    Jha, D., Smedsrud, P.H., Riegler, M.A., Halvorsen, P., De Lange, T., Johansen, D., Johansen, H.D.: Kvasir-seg: A segmented polyp dataset. In: MultiMedia modeling: 26th international conference, MMM 2020, Daejeon, South Korea, January 5–8, 2020, proceedings, part II 26. pp. 451...

    2020

  4. [12]

    (No Title) (2020)

    Jun, M., Cheng, G., Yixin, W., Xingle, A., Jiantao, G., Ziqi, Y., Minqing, Z., Xin, L., Xueyuan, D., Shucheng, C., et al.: Covid-19 ct lung and infection segmentation dataset. (No Title) (2020)

    2020

  5. [13]

    Radiology 304(3), 683–691 (2022)

    Kim, R.Y., Oke, J.L., Pickup, L.C., Munden, R.F., Dotson, T.L., Bellinger, C.R., Cohen, A., Simoff, M.J., Massion, P.P., Filippini, C., et al.: Artificial intelligence tool for assessment of indeterminate pulmonary nodules detected with ct. Radiology 304(3), 683–691 (2022)

    2022

  6. [14]

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

    Lin, T., Chen, Z., Yan, Z., Yu, W., Zheng, F.: Stable diffusion segmentation for biomedical images with single-step reverse process. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 656–666. Springer (2024)

    2024

  7. [15]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Liu, Y., Zhu, H., Liu, M., Yu, H., Chen, Z., Gao, J.: Rolling-unet: Revitalizing mlp’s ability to efficiently extract long-distance dependencies for medical image segmentation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 3819–3827 (2024)

    2024

  8. [16]

    arXiv preprint arXiv:2403.12036 (2024)

    Parmar, G., Park, T., Narasimhan, S., Zhu, J.Y.: One-step image translation with text-to-image models. arXiv preprint arXiv:2403.12036 (2024)

    2024 arXiv

  9. [17]

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

    Rahman, A., Valanarasu, J.M.J., Hacihaliloglu, I., Patel, V.M.: Ambiguous medical image segmentation using diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11536–11546 (2023)

    2023

  10. [18]

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

    Rahman, M.M., Munir, M., Marculescu, R.: Emcad: Efficient multi-scale convo- lutional attention decoding for medical image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11769–11779 (2024)

    2024

  11. [19]

    In: Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, Oc- tober 5-9, 2015, proceedings, part III 18

    Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomed- ical image segmentation. In: Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, Oc- tober 5-9, 2015, proceedings, part III 18....

    2015

  12. [20]

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

    Roy, S., Koehler, G., Ulrich, C., Baumgartner, M., Petersen, J., Isensee, F., Jaeger, P.F., Maier-Hein, K.H.: Mednext: transformer-driven scaling of convnets for medical image segmentation. In: International Conference on Medical Image Computing and Computer-Assisted Intervent...

    2023

  13. [21]

    In: International conference on medical image computing and computer-assisted intervention

    Ruan, J., Xie, M., Gao, J., Liu, T., Fu, Y.: Ege-unet: an efficient group enhanced unet for skin lesion segmentation. In: International conference on medical image computing and computer-assisted intervention. pp. 481–490. Springer (2023)

    2023

  14. [22]

    Frontiers in Medicine10, 1114673 (2023)

    Shi, L., Li, X., Hu, W., Chen, H., Chen, J., Fan, Z., Gao, M., Jing, Y., Lu, G., Ma, D., et al.: Ebhi-seg: A novel enteroscope biopsy histopathological hematoxylin and eosin image dataset for image segmentation tasks. Frontiers in Medicine10, 1114673 (2023)

    2023

  15. [23]

    International journal of computer assisted radiology and surgery9, 283–293 (2014)

    Silva, J., Histace, A., Romain, O., Dray, X., Granado, B.: Toward embedded detec- tion of polyps in wce images for early diagnosis of colorectal cancer. International journal of computer assisted radiology and surgery9, 283–293 (2014)

    2014

  16. [24]

    arXiv preprint arXiv:2010.02502 (2020) Title Suppressed Due to Excessive Length 11

    Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020) Title Suppressed Due to Excessive Length 11

    2020 arXiv

  17. [25]

    IEEE TMI35(2), 630–644 (2015)

    Tajbakhsh, N., Gurudu, S.R., Liang, J.: Automated polyp detection in colonoscopy videos using shape and context information. IEEE TMI35(2), 630–644 (2015)

    2015

  18. [26]

    Journal of healthcare engineering2017(1), 4037190 (2017)

    Vázquez, D., Bernal, J., Sánchez, F.J., Fernández-Esparrach, G., López, A.M., Romero, A., Drozdzal, M., Courville, A.: A benchmark for endoluminal scene segmentation of colonoscopy images. Journal of healthcare engineering2017(1), 4037190 (2017)

    2017

  19. [27]

    IEEE transactions on medical imaging23(7), 903–921 (2004)

    Warfield, S.K., Zou, K.H., Wells, W.M.: Simultaneous truth and performance level estimation (staple): an algorithm for the validation of image segmentation. IEEE transactions on medical imaging23(7), 903–921 (2004)

    2004

  20. [28]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Wu, J., Ji, W., Fu, H., Xu, M., Jin, Y., Xu, Y.: Medsegdiff-v2: Diffusion-based med- ical image segmentation with transformer. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 6030–6038 (2024)

    2024

  21. [29]

    arXiv preprint arXiv:2405.01002 (2024)

    Zhao, X., Pang, Y., Ji, W., Sheng, B., Zuo, J., Zhang, L., Lu, H.: Spider: A unified framework for context-dependent concept understanding. arXiv preprint arXiv:2405.01002 (2024)

    2024 arXiv

  22. [30]

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

    Zhu, L., Xue, Z., Jin, Z., Liu, X., He, J., Liu, Z., Yu, L.: Make-a-volume: Leveraging latent diffusion models for cross-modality 3d brain mri synthesis. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 592–601. Springer (2023)

    2023

  23. [31]

    In: Medical Image Computing and Computer Assisted Intervention–MICCAI 2019: 22nd International Conference, Shenzhen, China, October 13–17, 2019, Proceedings, Part VI 22

    Zhu, Z., Xia, Y., Xie, L., Fishman, E.K., Yuille, A.L.: Multi-scale coarse-to-fine segmentation for screening pancreatic ductal adenocarcinoma. In: Medical Image Computing and Computer Assisted Intervention–MICCAI 2019: 22nd International Conference, Shenzhen, China, October 1...

    2019

Pith tools

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