Latent Space Synergy: Text-Guided Data Augmentation for Direct Diffusion Biomedical Segmentation

Francesco Setti; Maham Nazir; Muhammad Aqeel; Zanxi Ruan

REVIEW 5 major objections 4 minor 30 references

Latent Space Synergy: Text-Guided Data Augmentation for Direct Diffusion Biomedical Segmentation

T0 review · 5 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read SynDiff: text-guided synthetic data plus single-step diffusion reaches 96.0% Dice on polyp segmentation.

desk verdict A plausible integration of known diffusion components for polyp segmentation, but the reported augmentation gain rests on synthetic masks whose provenance is not specified. read the letter →

arxiv 2507.15361 v1 pith:OO7VKP7O submitted 2025-07-21 eess.IV cs.AIcs.CV

Muhammad Aqeel , Maham Nazir , Zanxi Ruan , Francesco Setti This is my paper

classification eess.IVcs.AIcs.CV

keywords medicalimagesegmentationdiffusionmodelspolypdetectiontext-guidedsynthesisdataaugmentationsingle-stepinferencelatentCVC-ClinicDB

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

SynDiff claims that medical image segmentation can be improved by generating synthetic training images with text-guided inpainting and segmenting with a diffusion model that needs only a single inference step. On CVC-ClinicDB polyp segmentation, the combination reaches 96.0% Dice and 92.9% IoU, and the ablation attributes a +2.3 Dice gain to the synthetic augmentation. The paper argues that this shows controlled synthetic data can relieve data scarcity without shifting the distribution, and that binary segmentation masks are simple enough that iterative denoising can be replaced by direct latent estimation. A reader should care because the approach targets the two practical bottlenecks in clinical deployment: too few annotated images and too slow diffusion inference.

What carries the argument

The load-bearing mechanism is direct latent estimation: after the forward diffusion noising $z_t = \sqrt{\bar{\alpha}_t} z_0 + \sqrt{1-\bar{\alpha}_t} n$, the denoising U-Net predicts the noise $\tilde{n}=f(z_t,z_c)$, and the clean latent is recovered in one step as $\tilde{z}_0 = (z_t - \sqrt{1-\bar{\alpha}_t}\tilde{n})/\sqrt{\bar{\alpha}_t}$. Training minimizes both the noise loss and a direct latent loss $\mathcal{L} = \mathcal{L}_{\text{noise}} + \lambda \mathcal{L}_{\text{latent}}$ with $\lambda=1$. This dual supervision is what lets inference skip the denoising loop; the other pillar is offline text-guided inpainting, which uses SDXL to place a polyp inside a mask whose binary label is reused as ground truth.

What would settle it

Take the synthetic set described in Section 3.2 and measure the boundary distance between each generated polyp and its binary mask; if the average mismatch is large, the +2.3 Dice gain cannot be attributed to faithful labels. Alternatively, rerun the augmentation with masks drawn from the test split; if Dice still rises, the gain is not from the segmentation model learning genuine polyp structure.

Watch

Extended reading notes

Core claim

The central discovery is that text-conditioned SDXL inpainting can produce synthetic polyp images that, when used as additional training data, improve a diffusion-based segmenter more than geometric or GAN-based augmentation, while a direct latent estimation loss lets the same segmenter run in one forward pass. The paper reports that the full SynDiff pipeline reaches 96.0 ± 0.3 Dice and 92.9 ± 0.5 IoU on CVC-ClinicDB, that 100 synthetic samples (about 20% extra data) is the best augmentation budget, and that single-step inference at t = 50 is nearly indistinguishable from 50-step inference.

Load-bearing premise

The synthetic polyps must faithfully match the binary masks used as their training labels, and those masks must come only from the training split; the paper does not specify the split or verify contour alignment.

Editorial extensions

If this is right

Single-step inference at the fixed timestep $t=50$ matches the 50-step variant (96.0 vs 96.1 Dice), so the iterative denoising loop is unnecessary for binary segmentation masks.
Text-guided augmentation adds +2.3 Dice over the real-only baseline, while 200 synthetic samples reduce the gain, indicating an optimal augmentation budget near 20% extra data.
At the same 100-sample budget, text-guided synthesis beats traditional geometric augmentation by +1.2 Dice and GAN-based synthesis by +0.8 Dice.
Inference drops to 0.08 seconds per image, a 22–28× speedup over Diff-Trans and SDSeg, which is the paper's evidence that real-time clinical use is feasible.

Reading between the lines

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

Beyond the paper: because the synthetic masks are perfect binary labels, part of the measured +2.3 Dice could be a label-regularization effect rather than new anatomical diversity; a control that adds real images with identical masks would separate the two.
Beyond the paper: the claim that synthetic augmentation avoids distribution shift is tested only on CVC-ClinicDB; cross-dataset evaluation (train on CVC-ClinicDB, test on Kvasir-SEG) would reveal whether the synthetic diversity actually generalizes.
Beyond the paper: direct latent estimation exploits the discrete, near-binary structure of polyp masks; transferring SynDiff to multi-class or soft-label segmentation would likely require rebalancing the dual loss.
Beyond the paper: a natural test of the mechanism is to vary the number of distinct text prompts while holding the sample count fixed; if Dice tracks prompt diversity, semantic control is doing the work rather than sheer image count.

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.

Desk Editor's Note

read the letter

Straight to the point: SynDiff is a workmanlike integration of two known ideas—DiffBoost’s text-guided SDXL inpainting and SDSeg’s single-step latent estimation—applied to polyp segmentation. The headline number (96.0 Dice, 0.08s inference) is new for this combination, and the ablation in Table 2 gives real evidence that the text-guided augmentation adds something over traditional or GAN-based augmentation. The method section is clear, the math in Eq. 3 is a straightforward algebraic rearrangement and not circular, and the writing is honest about the single-dataset limitation.

The soft spots, in order of severity. First, the evaluation protocol is internally inconsistent. Section 4.1 says 488/62 split but also claims 5-fold cross-validation; Table 2’s baseline is described as “real data only” using the full 488 images, and no table states which protocol produced the reported numbers. That has to be fixed before anyone can trust the +2.3 Dice claim. Second, and more load-bearing: Section 3.2 says the inpainting masks are “derived from original dataset annotations” but never says they come only from the training split. If any test mask or test image was used to synthesize training data, the reported improvement is circular and the evaluation is invalid. The paper also doesn’t show any overlay verifying that the inpainted polyp actually aligns with the mask, so the synthetic labels are unverified. Third, the SDSeg comparison contradicts its own reference: the text says SDSeg “requires multiple denoising steps” but the cited paper [16] is titled “single-step reverse process.” That undercuts the claimed computational advantage and suggests the baseline numbers may not be from the cited method. Minor: the choice of 100 samples and t=50 is made on the same benchmark, so the results are positively selected, but the effect is small relative to the main flaw.

These are addressable. If the authors release code and data, specify the mask split, show overlay validation, and correct the baseline description, this could be a solid subfield contribution. As written, it should not be accepted without revision. I’d send it to peer review—the idea deserves referee time—but my own verdict is conditional.

Referee Report

5 major / 4 minor

Summary. The paper proposes SynDiff, a framework that combines offline text-guided synthetic data augmentation with a single-step diffusion-based segmentation model for polyp segmentation. Synthetic training images are generated with SDXL inpainting conditioned on clinical text prompts and binary masks; the segmentation model is a latent diffusion U-Net trained with both noise-prediction and direct-latent losses, and at inference it estimates the segmentation latent in one forward pass at a fixed timestep. On CVC-ClinicDB the paper reports 96.0% Dice and 92.9% IoU with 0.08 s inference time, and an ablation shows a +2.3 Dice gain from text-guided augmentation over a real-only baseline. The claims thus address both data scarcity and computational efficiency, but the current manuscript leaves several protocol and provenance questions unresolved.

Significance. If the evaluation is sound, the paper makes a practical contribution: a single-step diffusion segmentation method with a large measured speedup over iterative diffusion baselines, plus a text-controlled augmentation scheme that appears to improve both overlap and boundary metrics. The manuscript has clear strengths: the ablation in Tables 2 and 3 directly tests the augmentation contribution; Equation (3) is an algebraically correct rearrangement of the diffusion forward process; and the inclusion of HD95 and NSD metrics is appropriate for clinical segmentation. However, the load-bearing claims depend on details that are not currently specified: the evaluation protocol, the provenance of synthetic masks, the selection of hyperparameters, and the exact inference-time construction of the noisy latent. These issues are fixable within the manuscript's scope, but they must be resolved before the reported numbers can be accepted as stated.

major comments (5)

[Section 4.1] The evaluation protocol is stated inconsistently. Section 4.1 first says results are reported 'across 5-fold cross-validation' and then immediately defines '488 training and 62 test images.' Tables 1–3 report mean±std but do not state whether the numbers come from the fixed split or from cross-validation. Table 3 is described relative to 'our full training set of 488 real scans,' suggesting the fixed split was used; if so, the 5-fold claim is incorrect and the meaning of the ± values (multiple runs or folds) needs to be defined. Please specify the protocol used for each table, and if both protocols were used, report them separately.
[Section 3.2 / Section 4.2] The provenance of the synthetic masks is unspecified. The paper says binary masks are 'derived from original dataset annotations' (Section 4.2) and that these masks serve as ground truth for synthetic images (Section 3.2), but it never states that the masks and images used for inpainting are drawn exclusively from the 488 training images. If any test image or test annotation is used to generate synthetic training samples, the reported +2.3 Dice improvement and the headline 96.0 Dice would not measure generalization. Please state explicitly that only training-split images and masks are used, and provide overlay visualizations or a quantitative boundary-alignment check demonstrating that the inpainted polyp contours match the provided masks.
[Section 3.5 / Table 3] Several hyperparameters appear to be selected on the same benchmark used for the final accuracy claim. The fixed inference timestep t=50 is described as 'empirically determined' (Section 3.5), λ=1 in Eq. (6) is fixed without an ablation, and the choice of 100 synthetic samples is the optimum of Table 3 evaluated on CVC-ClinicDB. Unless a held-out validation split or nested selection procedure is used, the reported 96.0 Dice is a test-set-selected number. Please describe how t, λ, and the augmentation count were chosen and whether any selection was performed on the test partition.
[Table 4 / reference [16]] The computational comparison to SDSeg may be mis-stated. Table 4 reports that SDSeg requires 100 inference steps and 2.3 s, but the bibliography entry [16] is titled 'Stable diffusion segmentation for biomedical images with single-step reverse process,' which indicates that SDSeg is itself a single-step method. If that is the case, the claimed 22–28× speedup relative to SDSeg is not supported by the table. Please report the actual number of reverse steps and the measured runtime for each baseline, and reconcile the description of SDSeg with its published method.
[Section 3.5 / Section 4.2] The single-step inference procedure is under-specified. During training, z_t is obtained by forward-diffusing the ground-truth mask latent via Eq. (1), but at inference the mask is unknown. The text states only that 'Gaussian noise concatenated to image latents for direct mask estimation' is used (Section 4.2), without explaining whether z_t is pure Gaussian noise, a noise-corrupted placeholder, or something else, or how the fixed timestep t=50 is incorporated into the U-Net input. Please provide the exact inference-time construction of the noisy latent and the precise input to f(·); without this, the single-step claim cannot be reproduced.

minor comments (4)

[Abstract / Section 4.5] The abstract states that augmentation improves segmentation 'without distribution shift,' but Section 4.5 attributes the degradation at 200 synthetic samples to 'distribution shift.' Please reconcile or qualify this wording.
[Section 3.2] The source of the 'normal endoscopic image' used for inpainting is not defined; CVC-ClinicDB consists of polyp-containing frames, so it is unclear where the normal images come from. Please specify the image pool and how normal cases were identified.
[Table 2 / Table 3] The 'traditional augmentation' row in Tables 2 and 3 does not specify which geometric transformations were used or their hyperparameters, making the +1.2 Dice comparison with text-guided augmentation difficult to interpret. Please add these details.
[Equation (6)] The dual-supervision claim would be strengthened by ablating λ (e.g., latent-loss-only or a λ sweep); the current 'Noise loss only (λ=0)' row shows only one end of the spectrum.

Circularity Check

2 steps flagged · score 5.0 of 10

Reported 96.0 Dice is a test-set-selected maximum; the 'dual supervision' loss is a rescaling of the noise-prediction loss.

fitted input called prediction [Section 4.5, Table 3; also Section 3.5 (t=50)]
"The optimal performance is achieved with 100 synthetic samples (approximately 20% augmentation), where insufficient augmentation (20-50 samples) limits model exposure to variability, while excessive synthetic data (200 samples) introduces slight performance degradation due to distribution shift."

Table 3 sweeps synthetic-sample counts 0, 20, 50, 100, 200 and reports test-set Dice values 93.7, 94.1, 95.1, 96.0, 95.6. The paper then adopts the count with the highest test Dice (100) and reports 96.0 as the method's headline result. Thus the headline number is the maximum of a small test-set sweep, i.e., a hyperparameter fitted to the evaluation set and then presented as the method's performance, not an out-of-sample prediction. The same applies to t=50, said in Section 3.5 to be 'empirically determined for optimal quality-efficiency balance' on the same dataset. The +2.3 Dice augmentation gain is therefore partly a selected optimum, although the direction of the improvement is not forced.
self definitional [Section 3.3, Eqs. 3-6]
"The key innovation lies in directly estimating the clean latent z0 from the noisy latent zt in a single step through: ˜z0 = 1/√¯αt (zt − √(1 − ¯αt) ˜n) (3) This enables a dual supervision strategy where we minimize both noise prediction error: Lnoise = ||n − ˜n||1 (4) and direct latent estimation error: Llatent = ||z0 − ˜z0||1 (5)"

Substituting Eq. 3 into Eq. 5 gives z0 − ˜z0 = √((1−¯αt)/¯αt) (˜n − n), so Llatent = √((1−¯αt)/¯αt) · Lnoise for every training sample. The 'direct latent estimation' objective is therefore a timestep-dependent rescaling of the noise-prediction objective; the two losses carry identical information. The claimed simultaneous learning of 'denoising and direct prediction capabilities' is self-definitional: the second capability is defined in terms of the first. The λ ablation in Table 2 ('Noise loss only (λ=0)' versus the full loss) consequently compares two loss weightings rather than two distinct supervision signals, so the stated benefit of 'dual supervision' is not an independent mechanism.

full rationale

The central augmentation claim is not wholly circular: SynDiff is evaluated on real CVC-ClinicDB test images against external baselines, and the +2.3 Dice gain over the real-only baseline is an empirical difference, not a formal consequence of the training objective. However, two steps are partly self-referential. First, the number of synthetic samples (100) and the inference timestep (t=50) are selected using the same test set that produces the headline numbers; Table 3 is a sweep over 0/20/50/100/200 samples, so 96.0 Dice is the maximum of a test-set sweep, i.e., a fitted hyperparameter reported as the method's performance rather than an out-of-sample prediction. Second, the claimed 'dual supervision' loss is redundant by the paper's own equations: Eq. 3 makes the direct latent estimate an affine function of the predicted noise, so Llatent equals a per-timestep rescaling of Lnoise; the ablation therefore compares loss weightings, not two independent supervision signals. Self-citations [1-4,11] are contextual and not load-bearing; the single-step reverse-process idea is attributed to the independent MICCAI paper [16]. The manuscript's failure to state whether the inpainting masks come only from the 488 training images is a potential leakage risk that would make the evaluation circular, but the text provides no reduction that would let us prove such leakage, so it is recorded as a correctness risk rather than a scored circular step.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The paper does not introduce new physical or conceptual entities. The free parameters listed are design choices selected on the evaluation dataset without a documented validation protocol. The axioms reflect assumptions about synthetic label correctness, data provenance, and feature transfer that are not independently verified.

free parameters (4)

Inference timestep t = 50
Empirically selected for quality-efficiency balance in Section 3.5; inference outcome depends on this choice and no validation set is described.
Augmentation count = 100
Table 3 shows a peak at 100 synthetic images among 20, 50, 100, and 200; choosing the best value on the evaluation set is post hoc.
Dual supervision weight lambda = 1
Set to 1 in Section 3.3 without ablation; it controls the balance between noise prediction and latent estimation losses.
Hand-crafted text prompts = 50 prompts for 100 samples
The prompt set is not released, so its composition is an uncontrolled modeling choice that shapes the synthetic data.

assumptions (6)

standard math DDPM forward noise process (Eq. 1) and the inversion formula (Eq. 3) are valid.
Standard diffusion modeling from Ho et al. [13]; accepted as background.
domain assumption SDXL inpainted polyps conform to the binary masks used as labels.
Section 3.2 treats mask ms as ground truth for the generated image is; no verification of boundary alignment is provided.
domain assumption Normal endoscopic frames are available for inpainting.
Section 3.2 samples 'normal endoscopic images' from available normal cases, but CVC-ClinicDB is not described as containing a normal-case source.
domain assumption Synthetic masks are drawn only from training split annotations.
Section 3.2 says masks are derived from original dataset annotations but does not specify the split; leakage would invalidate the evaluation.
ad hoc to paper Single-step inference at fixed timestep t=50 is valid.
Section 3.5 fixes t=50 empirically rather than deriving it from the model or a validation protocol.
domain assumption Pre-trained SDXL and autoencoder features transfer to endoscopy.
Sections 3.2 and 4.2 use frozen pre-trained weights; transfer is assumed rather than demonstrated.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Latent Space Synergy: Text-Guided Data Augmentation for Direct Diffusion Biomedical Segmentation." pith.science (2026). https://pith.science/paper/OO7VKP7O

@misc{pith2026250715361,
  author       = {Pith},
  title        = {Pith review of: Latent Space Synergy: Text-Guided Data Augmentation for Direct Diffusion Biomedical Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OO7VKP7O}},
  note         = {Machine review of arXiv:2507.15361}
}

read the original abstract

Medical image segmentation suffers from data scarcity, particularly in polyp detection where annotation requires specialized expertise. We present SynDiff, a framework combining text-guided synthetic data generation with efficient diffusion-based segmentation. Our approach employs latent diffusion models to generate clinically realistic synthetic polyps through text-conditioned inpainting, augmenting limited training data with semantically diverse samples. Unlike traditional diffusion methods requiring iterative denoising, we introduce direct latent estimation enabling single-step inference with T x computational speedup. On CVC-ClinicDB, SynDiff achieves 96.0% Dice and 92.9% IoU while maintaining real-time capability suitable for clinical deployment. The framework demonstrates that controlled synthetic augmentation improves segmentation robustness without distribution shift. SynDiff bridges the gap between data-hungry deep learning models and clinical constraints, offering an efficient solution for deployment in resourcelimited medical settings.

Figures

Figures reproduced from arXiv: 2507.15361 by the authors.

**Figure 1.** Overview of the SynDiff framework. (Left) Offline text-guided data generation: Stable Diffusion XL (SDXL) inpainting takes a normal endoscopic image, clinical text prompt, and binary mask to generate synthetic polyp images with corresponding ground truth masks. (Right) Single-step segmentation pipeline: input images are encoded through a trainable vision encoder, processed by a denoising U-Net for direct latent es… view at source ↗

**Figure 2.** Synthetically generated endoscopic images with corresponding binary segmentation masks. The first and third rows display synthetic colonoscopy images created through our text-guided generation approach, while the second and fourth rows present their corresponding binary masks. White regions in the masks indicate the synthetic anomalous tissue generated through the inpainting process. This augmented dataset enhances… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

30 extracted references · 14 canonical work pages

[16]

In: Medical Image Computing andComputerAssistedIntervention–MICCAI2024.pp.656–666.SpringerNature Switzerland, Cham (2024)

Lin, T., Chen, Z., Yan, Z., Yu, W., Zheng, F.: Stable diffusion segmentation for biomedical images with single-step reverse process. In: Medical Image Computing andComputerAssistedIntervention–MICCAI2024.pp.656–666.SpringerNature Switzerland, Cham (2024)

work page 2024
[1]

In: European Confer- ence on Computer Vision

Aqeel, M., Sharifi, S., Cristani, M., Setti, F.: Meta learning-driven iterative refine- ment for robust anomaly detection in industrial inspection. In: European Confer- ence on Computer Vision. pp. 445–460. Springer (2024)

work page 2024
[2]

In: International Conference on Deep Learning Theory and Applications (2024)

Aqeel, M., Sharifi, S., Cristani, M., Setti, F.: Self-supervised learning for robust surface defect detection. In: International Conference on Deep Learning Theory and Applications (2024)

work page 2024
[3]

In: International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (2025)

Aqeel, M., Sharifi, S., Cristani, M., Setti, F.: Self-supervised iterative refinement for anomaly detection in industrial quality control. In: International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (2025)

work page 2025
[4]

In: Accepted to Proceedings of the ieee/cvf international conference on computer vision (2025)

Aqeel, M., Sharifi, S., Cristani, M., Setti, F.: Towards real unsupervised anomaly detection via confident meta-learning. In: Accepted to Proceedings of the ieee/cvf international conference on computer vision (2025)

work page 2025
[5]

arXiv preprint arXiv:2112.03126 (2021)

Baranchuk, D., Rubachev, I., Voynov, A., Khrulkov, V., Babenko, A.: Label-efficient semantic segmentation with diffusion models. arXiv preprint arXiv:2112.03126 (2021)

arXiv 2021
[6]

saliency maps from physicians

Bernal, J., Sánchez, F.J., Fernández-Esparrach, G., Gil, D., Rodríguez, C., Vilar- iño, F.: Wm-dova maps for accurate polyp highlighting in colonoscopy: Validation vs. saliency maps from physicians. Computerized medical imaging and graphics 43, 99–111 (2015)

2015
[7]

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:Transformersmakestrongencodersformedicalimagesegmentation. arXiv preprint arXiv:2102.04306 (2021)

arXiv 2021

Show all 30 references

[8]

In: Machine learning for healthcare conference

Choi, E., Biswal, S., Malin, B., Duke, J., Stewart, W.F., Sun, J.: Generating multi- label discrete patient records using generative adversarial networks. In: Machine learning for healthcare conference. pp. 286–305. PMLR (2017)

2017
[9]

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

Chowdary, G.J., Yin, Z.: Diffusion transformer u-net for medical image segmen- tation. In: International Conference on Medical Image Computing and Computer- Assisted Intervention. pp. 622–631. Springer (2023)

2023
[10]

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

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: International confer- ence on medical image computing and computer-assisted intervention. pp. 263–273. Springer (2020)

2020
[11]

In: International Conference on Content-Based Multi- media Indexing (CBMI) (2024)

Girella, F., Liu, Z., Fummi, F., Setti, F., Cristani, M., Capogrosso, L.: Leveraging Latent Diffusion Models for Training-Free in-Distribution Data Augmentation for Surface Defect Detection. In: International Conference on Content-Based Multi- media Indexing (CBMI) (2024)

2024
[12]

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
[13]

Advances in neural information processing systems33, 6840–6851 (2020)

Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems33, 6840–6851 (2020)

2020
[14]

Nature methods18(2), 203–211 (2021)

Isensee, F., Jaeger, P.F., Kohl, S.A., Petersen, J., Maier-Hein, K.H.: nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods18(2), 203–211 (2021)

2021
[15]

In: International con- ference on multimedia modeling

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: International con- ference on multimedia modeling. pp. 451–462. Springer (2019) 12 M.Aqeel et al

2019
[17]

In: International conference on machine learning

Nichol, A.Q., Dhariwal, P.: Improved denoising diffusion probabilistic models. In: International conference on machine learning. pp. 8162–8171. PMLR (2021)

2021
[18]

Medical image analysis59, 101570 (2020)

Orlando, J.I., Fu, H., Breda, J.B., Van Keer, K., Bathula, D.R., Diaz-Pinto, A., Fang, R., Heng, P.A., Kim, J., Lee, J., et al.: Refuge challenge: A unified frame- work for evaluating automated methods for glaucoma assessment from fundus pho- tographs. Medical image analysis59...

2020
[19]

Peebles,W.,Xie,S.:Scalablediffusionmodelswithtransformers.In:Proceedingsof the IEEE/CVF international conference on computer vision. pp. 4195–4205 (2023)

2023
[20]

arXiv preprint arXiv:2307.01952 (2023)

Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023)

2023 arXiv
[21]

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

Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684–10695 (2022)

2022
[22]

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
[23]

IEEE Transactions on Biomedical Engineering70(4), 1330–1339 (2022)

Sharma, P., Gautam, A., Maji, P., Pachori, R.B., Balabantaray, B.K.: Li-segpnet: Encoder-decoder mode lightweight segmentation network for colorectal polyps analysis. IEEE Transactions on Biomedical Engineering70(4), 1330–1339 (2022)

2022
[24]

In: 2022 IEEE 24th international workshop on multimedia signal processing (MMSP)

Shi, W., Xu, J., Gao, P.: Ssformer: A lightweight transformer for semantic segmen- tation. In: 2022 IEEE 24th international workshop on multimedia signal processing (MMSP). pp. 1–5. IEEE (2022)

2022
[25]

Song, Y., Dhariwal, P., Chen, M., Sutskever, I.: Consistency models (2023)

2023
[26]

In: Medical Image Computing and Computer Assisted Intervention–MICCAI 2021: 24th International Conference, Strasbourg, France, September 27–October 1, 2021, Proceedings, Part I 24

Wei, J., Hu, Y., Zhang, R., Li, Z., Zhou, S.K., Cui, S.: Shallow attention network for polyp segmentation. In: Medical Image Computing and Computer Assisted Intervention–MICCAI 2021: 24th International Conference, Strasbourg, France, September 27–October 1, 2021, Proceedings, ...

2021
[27]

In: Medical Imaging with Deep Learning (2024)

Wu, J., Fu, R., Fang, H., Zhang, Y., Yang, Y., Xiong, H., Liu, H., Xu, Y.: Med- segdiff: Medical image segmentation with diffusion probabilistic model. In: Medical Imaging with Deep Learning (2024)

2024
[28]

In: Proceedings of the AAAI conference on artificial intelligence (2024)

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 (2024)

2024
[29]

arXiv preprint arXiv:2303.10326 (2023)

Xing, Z., Wan, L., Fu, H., Yang, G., Zhu, L.: Diff-unet: A diffusion embedded network for volumetric segmentation. arXiv preprint arXiv:2303.10326 (2023)

2023 arXiv
[30]

IEEE Transactions on Medical Imaging (2024)

Zhang, Z., Yao, L., Wang, B., Jha, D., Durak, G., Keles, E., Medetalibeyoglu, A., Bagci, U.: Diffboost: Enhancing medical image segmentation via text-guided diffusion model. IEEE Transactions on Medical Imaging (2024)

2024

Pith tools

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

[1] [16]

In: Medical Image Computing andComputerAssistedIntervention–MICCAI2024.pp.656–666.SpringerNature Switzerland, Cham (2024)

Lin, T., Chen, Z., Yan, Z., Yu, W., Zheng, F.: Stable diffusion segmentation for biomedical images with single-step reverse process. In: Medical Image Computing andComputerAssistedIntervention–MICCAI2024.pp.656–666.SpringerNature Switzerland, Cham (2024)

work page 2024

[2] [1]

In: European Confer- ence on Computer Vision

Aqeel, M., Sharifi, S., Cristani, M., Setti, F.: Meta learning-driven iterative refine- ment for robust anomaly detection in industrial inspection. In: European Confer- ence on Computer Vision. pp. 445–460. Springer (2024)

work page 2024

[3] [2]

In: International Conference on Deep Learning Theory and Applications (2024)

Aqeel, M., Sharifi, S., Cristani, M., Setti, F.: Self-supervised learning for robust surface defect detection. In: International Conference on Deep Learning Theory and Applications (2024)

work page 2024

[4] [3]

In: International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (2025)

Aqeel, M., Sharifi, S., Cristani, M., Setti, F.: Self-supervised iterative refinement for anomaly detection in industrial quality control. In: International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (2025)

work page 2025

[5] [4]

In: Accepted to Proceedings of the ieee/cvf international conference on computer vision (2025)

Aqeel, M., Sharifi, S., Cristani, M., Setti, F.: Towards real unsupervised anomaly detection via confident meta-learning. In: Accepted to Proceedings of the ieee/cvf international conference on computer vision (2025)

work page 2025

[6] [5]

arXiv preprint arXiv:2112.03126 (2021)

Baranchuk, D., Rubachev, I., Voynov, A., Khrulkov, V., Babenko, A.: Label-efficient semantic segmentation with diffusion models. arXiv preprint arXiv:2112.03126 (2021)

arXiv 2021

[7] [6]

saliency maps from physicians

Bernal, J., Sánchez, F.J., Fernández-Esparrach, G., Gil, D., Rodríguez, C., Vilar- iño, F.: Wm-dova maps for accurate polyp highlighting in colonoscopy: Validation vs. saliency maps from physicians. Computerized medical imaging and graphics 43, 99–111 (2015)

2015

[8] [7]

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:Transformersmakestrongencodersformedicalimagesegmentation. arXiv preprint arXiv:2102.04306 (2021)

arXiv 2021