Pith. sign in

REVIEW 4 major objections 4 minor 42 references

Diffusion-Based Data Augmentation for Medical Image Segmentation

T0 review · 4 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A text-and-mask guided diffusion pipeline with a spatial quality gate synthesizes training lesions that lift medical segmentation Dice to 96.4%.

desk verdict Sensible diffusion-augmentation pipeline, but the central ablation appears to compare against the wrong baseline, so the headline Dice gains are not yet supported. read the letter →

arxiv 2508.17844 v1 pith:57K4WPDO submitted 2025-08-25 cs.CV cs.LG

Maham Nazir , Muhammad Aqeel , Francesco Setti This is my paper
classification cs.CVcs.LG
keywords diffusionmodelsdataaugmentationmedicalimagesegmentationtext-guidedinpaintinglatentspacevalidationpolypglaucomadetection
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

The paper proposes DiffAug, a data-augmentation pipeline that uses text-guided latent diffusion inpainting to synthesize medical abnormalities—polyps, flat lesions, glaucomatous cups—onto healthy images, then passes each synthetic image through a segmentation network that checks whether the generated abnormality is located where the prompt mask says it should be. The authors claim that this generation-plus-validation loop produces synthetic training data diverse and anatomically plausible enough to improve downstream segmentation, and they report state-of-the-art Dice coefficients on CVC-ClinicDB (96.4%), Kvasir-SEG (95.6%), and REFUGE2 (90.2%), with up to 28% relative reductions in false negatives for small polyps and flat lesions compared to a U-Net baseline. The central idea is that text gives variety, masks give placement, and a validation gate—not manual review—ensures quality. If the claim holds, scarce-pathology segmentation can be improved without collecting more annotated abnormal cases, and text prompts derived from medical terminology replace manual annotation of synthetic data.

What carries the argument

The load-bearing mechanism is the DiffAug generate-and-gate loop. Generation uses a pre-trained SDXL inpainting model in a compressed latent space: a text encoder supplies semantic prompts such as 'small sessile polyp' or 'inferior rim thinning', and a mask from an existing annotation fixes the location, while a loss term preserves the surrounding healthy tissue. Validation is carried out by a segmentation network trained to denoise directly in the latent space, so it can predict a mask in a single step instead of iterating; the pipeline accepts a synthetic image only when the predicted mask's IoU with the intended mask exceeds 0.7. This gate is what turns raw generations into training data,

What would settle it

Generate two batches of synthetic images that both pass the IoU > 0.7 gate, one with correct medical text prompts and one with scrambled prompts; if downstream Dice does not drop for the scrambled batch, text guidance is not responsible for the gain. Alternatively, lower the gate to IoU = 0.5 and retrain: if Dice stays at the same level, the threshold is not the active ingredient.

Watch

Extended reading notes

Core claim

DiffAug's central claim is that a latent diffusion model fine-tuned for inpainting can turn normal medical images into training-relevant abnormal ones when conditioned on a medical text snippet and a binary mask, and that this synthetic data becomes useful for segmentation only after a latent-space segmentation network filters it by spatial agreement (IoU > 0.7). The authors show this on three benchmarks: adding up to 3× synthetic samples raises polyp and optic-disc/cup segmentation Dice to 96.4%, 95.6%, and 90.2%, respectively, and the largest clinical gains are for small polyps (<5 mm), flat lesions, and early glaucoma (cup-to-disc ratio 0.5–0.6), where false-negative rates drop relative t

Load-bearing premise

Everything rests on the assumption that masked text-guided SDXL inpainting produces synthetic abnormalities anatomically plausible enough that IoU > 0.7 agreement with a latent segmentation network selects samples that genuinely improve a separate downstream segmentor, rather than merely matching that validator's inductive bias.

Editorial extensions

If this is right

  • Training on roughly 3× synthetic data relative to real images is where Dice gains saturate; beyond that, added synthetic volume does not help.
  • About 20–25% of initial generations fail the IoU > 0.7 gate, mostly for subtle flat lesions, so the pipeline trades generation diversity for spatial accuracy.
  • During deployment only the segmentation network is used—12.5 samples/s—so the expensive diffusion step is an offline preprocessing cost, not a runtime cost.
  • The ablation without synthetic data (−8.1 Dice) suggests most of the measured gain is attributable to the synthetic augmentation itself, not to the validator or prompt design alone.

Reading between the lines

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

  • Because the acceptance rule is agreement with the validating segmentor, a natural extension is to check whether accepted samples also improve a downstream segmentor with a different architecture; the paper reports gains on its own segmentation backbone, so cross-architecture transfer remains an open test.
  • The same generate-and-gate recipe could transfer to other lesion types with small or flat presentations—such as Barrett's esophagus or early diabetic retinopathy—by writing new medical-text prompt banks; the paper does not test those domains.
  • Since rejected samples have clearly worse FID and LPIPS, the IoU gate is implicitly filtering for visual realism as well as localization; measuring how much of the gain comes from realism filtering versus spatial accuracy would isolate the gate's role.
  • The empirical threshold sweep (0.5–0.9) reported in the paper suggests the gate has a tunable precision–diversity trade-off; future work could adapt θquality per abnormality type rather than globally.
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 / 4 minor

Summary. The paper proposes DiffAug, a text-guided latent diffusion inpainting framework that synthesizes abnormal medical images by conditioning SDXL on medical text prompts and spatial masks derived from existing annotations. Generated images are filtered by a latent-space segmentation network through an IoU threshold (θ=0.7), and the accepted synthetic pairs are added to the real training data for a downstream diffusion-based segmentation network. The authors report state-of-the-art Dice scores of 96.4% on CVC-ClinicDB, 95.6% on Kvasir-SEG, and 90.2% on REFUGE2, with false-negative reductions up to 28% for small polyps and flat lesions. The paper includes component ablations, data-volume scaling experiments, and efficiency measurements.

Significance. If the central claim holds, the framework is practically valuable: it leverages a pretrained text-to-image model, requires no manual annotation of new abnormalities, filters samples automatically, and runs the downstream segmentor at 12.5 samples/s on a single GPU. The pipeline is presented with enough detail (Eqs. 1–14, Section 4.3) that the components are identifiable, and the ablation structure is a sensible attempt to attribute gains. However, the empirical support currently has load-bearing confounds: the no-synthetic control appears to be a different architecture, the main comparison does not match training-set size, and the quality gate may be circular with respect to the downstream segmentor. These issues need to be resolved before the headline claims can be accepted.

major comments (4)
  1. [§6.1, Table 6] The 'w/o Synthetic Data' row reports Dice 88.3±1.2 / IoU 81.4±1.8, numerically identical to the U-Net row in Table 2 and to the 'Baseline (no augmentation)' row in Table 7. Section 4.3 defines the DiffAug segmentor as a latent-space SD U-Net trained with a diffusion objective and single-step inference. If the no-synthetic row is the vanilla U-Net rather than this segmentor trained on real data only, the ablation does not isolate the augmentation effect: the 8.1-point gain could be due to architecture or training recipe. Please name the network used for each row and add an architecture-matched no-synthetic control.
  2. [§5.1, Table 2] It is not stated how many training images each row in Table 2 uses. Table 7 indicates that DiffAug full uses 488 real + 1464 synthetic images (3×), while the baselines likely use only 488 real images. The comparison therefore does not match training-set size, and part of the reported 8–10% gain may reflect additional data volume rather than the synthetic generation/validation mechanism. Please report the exact training data for every row or include a DiffAug variant trained with an equivalent amount of additional real data.
  3. [§3.3, Eq. (14)] The quality gate accepts samples with IoU > 0.7 between the intended mask and the prediction of the latent segmentation network S_phi. The downstream segmentor in §4.3 is built on the same SD U-Net architecture with the same single-step latent estimation. This is a self-selection mechanism: generated samples that challenge S_phi's inductive bias are discarded, so training on accepted samples may improve S_phi-type models without demonstrating generalizable anatomical plausibility. Please validate with an independent gate (e.g., a different architecture or a clinical metric) or evaluate downstream with a segmentor not used in the gate.
  4. [§5.1] The statement that the 0.6–0.8 percentage-point improvements over SDSeg are 'statistically significant (p < 0.05)' is unsupported: no test name, number of runs, or multiple-comparison correction is given. With overlapping standard deviations, this claim needs a paired test across folds or a bootstrap with a defined null. Please add the test details or soften the claim.
minor comments (4)
  1. [§3.3 vs. Table 5] The acceptance rate at θ=0.7 is reported as 65% in Section 3.3 but 78.3% in Table 5. Please reconcile these numbers.
  2. [§4.3, Eq. (7)] The mask M is defined in latent resolution {0,1}^{h×w}, while the segmentation masks used for training are presumably at image resolution. Please clarify how masks are resized/downsampled for the latent-space operations.
  3. [Table 3] The column 'Relative FNR Reduction (%)' appears to be relative to U-Net, but the text in §5.2 uses SDSeg as the denominator (e.g., 2.7/25.9). Please state the reference baseline explicitly for this column.
  4. [Global] Minor language issues: 'We present' is capitalized mid-sentence in the Introduction, 'These models shown' is missing a verb, and 'V AEs' has inconsistent spacing. A light copyedit would improve readability.

Circularity Check

2 steps flagged · score 6.0 of 10

The claimed gain from synthetic data is supported by an ablation row numerically identical to the vanilla U-Net baseline, and the quality gate filters samples using the same segmentation architecture that is later evaluated.

  1. renaming known result [Section 6.1, Table 6 vs. Section 5.1, Table 2]
    "Most significantly, training without synthetic augmentation results in an 8.1% Dice decrease and increases FNR from 23.2% to 31.7%, demonstrating the clear value of our data generation approach... Table 6: 'w/o Synthetic Data 88.3 ± 1.2 81.4 ± 1.8 31.7 12.5'; Table 2: 'U-Net [28] 88.3 ±1.2 / 81.4±1.8' with FNR 31.7."

    The control row for the central ablation is numerically identical to the vanilla U-Net baseline (Table 2: Dice 88.3, IoU 81.4, FNR 31.7; Table 6: Dice 88.3, IoU 81.4, FNR 31.7). The paper presents this as 'training without synthetic augmentation', but it is the traditional U-Net baseline, not the DiffAug segmentor (SD U-Net with latent single-step inference) trained on real data alone. The 8.1-point gain therefore conflates the architecture/training-recipe change with the synthetic-data contribution; the headline causal claim reduces to re-labeling the existing baseline as an ablation.

  2. self definitional [Section 3.3, Eq. (14); Section 4.3; Section 5.3]
    "IoU(i,j) = IoU(Mi, D( ˆM (i,j) latent )) ... Only generations exceeding a quality threshold θquality = 0.7 are retained... The validated synthetic pairs ... are combined with real training data for model training. Segmentation Network. Built upon the Stable Diffusion U-Net architecture... our network achieves 12.5 samples/second inference speed using single-step latent estimation."

    The acceptance criterion for synthetic data is the IoU between the intended mask and the prediction of the latent-space SD U-Net Sϕ. The same architecture (SD U-Net, single-step latent estimation) is the downstream segmentor whose Dice/IoU/FNR are reported. Hence accepted synthetic samples are by construction those that this architecture already segments with IoU>0.7; the downstream evaluation is performed with the same model family that defined the filter. The quality gate therefore selects for the segmentor's inductive bias rather than providing an independent measure of synthetic-data value, making part of the observed improvement a self-selection artifact.

full rationale

The diffusion-generation equations and the segmentation-training objective are internally consistent; there is no equation-level circularity in the generative model itself, and the external benchmarks are legitimate. However, the paper's central causal claim—that synthetic data improves segmentation—rests on two circular or near-circular controls. First, the ablation 'w/o Synthetic Data' in Table 6 has exactly the same Dice/IoU/FNR as the vanilla U-Net baseline in Table 2, so it is not an architecture-matched no-synthetic control; the 8.1% gain credited to synthetic data is actually a comparison between two different architectures/training recipes. Second, the quality gate that selects synthetic training samples uses the same latent-space SD U-Net architecture that is later evaluated as the downstream segmentor; samples are accepted precisely when this architecture can already localize them well, creating a self-selection bias that inflates the apparent benefit of the augmentation. The self-citation [1] is not load-bearing in the text, and the core methodological machinery is not imported from the authors' prior work, so the paper is not wholly derivative. But these two issues together make the headline improvements partly forced by the experimental setup rather than by independent evidence, warranting a score of 6.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The framework rests on unstated transfer assumptions and a hand-tuned quality gate. No new physical entities are introduced. Loss weights in Eqs. 7 and 12 are not reported, leaving hidden free parameters.

free parameters (4)
  • quality threshold θquality = 0.7
    Selected empirically from 0.5-0.9 by balancing acceptance rate (65%) and downstream performance (Section 3.3).
  • classifier-free guidance scale s = 7.5
    Chosen for generation fidelity/diversity balance (Section 4.4).
  • loss weights λpreserve, λlatent, λdice = not reported
    Equations 7 and 12 use unstated weights, which are free parameters affecting training.
  • number of denoising steps = 50
    Set in Section 4.4; tradeoff between speed and quality.
assumptions (5)
  • standard math Diffusion forward process and reparameterization (Eqs. 2-3) hold as standard.
    Standard DDPM formulations used without proof.
  • domain assumption SDXL VAE and text encoder pretrained on natural images transfer meaningfully to endoscopy and fundus images.
    Sections 3.1 and 4.3 assume this without adaptation or validation.
  • domain assumption Inpainting on normal images with masks from existing annotations yields anatomically plausible abnormalities.
    Section 3.1 relies on this for all generated samples.
  • ad hoc to paper IoU between intended mask and latent-segmentation prediction is a sufficient proxy for clinical plausibility.
    Section 3.3 defines quality entirely through this metric.
  • domain assumption Normal regions can be obtained by excluding annotated abnormalities, and these form representative healthy backgrounds.
    Section 4.1 states this extraction process without verification.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Diffusion-Based Data Augmentation for Medical Image Segmentation." pith.science (2026). https://pith.science/paper/57K4WPDO

@misc{pith2026250817844,
  author       = {Pith},
  title        = {Pith review of: Diffusion-Based Data Augmentation for Medical Image Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/57K4WPDO}},
  note         = {Machine review of arXiv:2508.17844}
}
read the original abstract

Medical image segmentation models struggle with rare abnormalities due to scarce annotated pathological data. We propose DiffAug a novel framework that combines textguided diffusion-based generation with automatic segmentation validation to address this challenge. Our proposed approach uses latent diffusion models conditioned on medical text descriptions and spatial masks to synthesize abnormalities via inpainting on normal images. Generated samples undergo dynamic quality validation through a latentspace segmentation network that ensures accurate localization while enabling single-step inference. The text prompts, derived from medical literature, guide the generation of diverse abnormality types without requiring manual annotation. Our validation mechanism filters synthetic samples based on spatial accuracy, maintaining quality while operating efficiently through direct latent estimation. Evaluated on three medical imaging benchmarks (CVC-ClinicDB, Kvasir-SEG, REFUGE2), our framework achieves state-of-the-art performance with 8-10% Dice improvements over baselines and reduces false negative rates by up to 28% for challenging cases like small polyps and flat lesions critical for early detection in screening applications.

Figures

Figures reproduced from arXiv: 2508.17844 by the authors.

Figure 1
Figure 1. Overview of our integrated generation-validation pipeline. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Examples of synthetic abnormalities generated by our [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

42 extracted references · 34 canonical work pages

  1. [1]

    Latent space synergy: Text-guided data aug- mentation for direct diffusion biomedical segmentation

    Muhammad Aqeel, Maham Nazir, Zanxi Ruan, and Francesco Setti. Latent space synergy: Text-guided data aug- mentation for direct diffusion biomedical segmentation. In Proceedings of International Conference on Image Analysis and Processing, 2025. 2

    work page 2025

  2. [2]

    Towards real unsupervised anomaly de- tection via confident meta-learning

    Muhammad Aqeel, Shakiba Sharifi, Marco Cristani, and Francesco Setti. Towards real unsupervised anomaly de- tection via confident meta-learning. In Proceedings of the ieee/cvf international conference on computer vision , 2025. 2

    work page 2025

  3. [3]

    Wasserstein generative adversarial networks

    Martin Arjovsky, Soumith Chintala, and L ´eon Bottou. Wasserstein generative adversarial networks. In Interna- tional conference on machine learning , pages 214–223. PMLR, 2017. 1

    work page 2017

  4. [4]

    Autoencoders for unsuper- vised anomaly segmentation in brain mr images: a compara- tive study

    Christoph Baur, Stefan Denner, Benedikt Wiestler, Nassir Navab, and Shadi Albarqouni. Autoencoders for unsuper- vised anomaly segmentation in brain mr images: a compara- tive study. Medical image analysis, 69:101952, 2021. 2

    work page 2021

  5. [5]

    Wm-dova maps for accurate polyp highlighting in colonoscopy: Validation vs

    Jorge Bernal, F Javier S ´anchez, Gloria Fern ´andez- Esparrach, Debora Gil, Cristina Rodr ´ıguez, and Fernando Vilari˜no. Wm-dova maps for accurate polyp highlighting in colonoscopy: Validation vs. saliency maps from physi- cians. Computerized medical imaging and graphics, 43:99– 111, 2015. 5, 6

    work page 2015

  6. [6]

    Clinical-grade computational pathology using weakly supervised deep learning on whole slide images

    Gabriele Campanella, Matthew G Hanna, Luke Geneslaw, Allen Miraflor, Vitor Werneck Krauss Silva, Klaus J Busam, Edi Brogi, Victor E Reuter, David S Klimstra, and Thomas J Fuchs. Clinical-grade computational pathology using weakly supervised deep learning on whole slide images. Nature medicine, 25(8):1301–1309, 2019. 1

    work page 2019

  7. [7]

    Roentgen: vision-language foundation model for chest x-ray generation

    Pierre Chambon, Christian Bluethgen, Jean-Benoit Del- brouck, Rogier Van der Sluijs, Małgorzata Połacin, Juan Manuel Zambrano Chaves, Tanishq Mathew Abraham, Shiv- anshu Purohit, Curtis P Langlotz, and Akshay Chaudhari. Roentgen: vision-language foundation model for chest x-ray generation. arXiv preprint arXiv:2211.12737, 2022. 2

    arXiv 2022

  8. [8]

    Transunet: Transformers make strong encoders for medi- cal image segmentation

    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 medi- cal image segmentation. arXiv preprint arXiv:2102.04306,

    arXiv

Show all 42 references
  1. [9]

    Diffusion trans- former u-net for medical image segmentation

    G Jignesh Chowdary and Zhaozheng Yin. Diffusion trans- former u-net for medical image segmentation. In In- ternational conference on medical image computing and computer-assisted intervention , pages 622–631. Springer,

  2. [10]

    Diffusion models beat gans on image synthesis

    Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural informa- tion processing systems, 34:8780–8794, 2021. 1, 2

    2021

  3. [11]

    Deep learning-enabled medical com- puter vision

    Andre Esteva, Katherine Chou, Serena Yeung, Nikhil Naik, Ali Madani, Ali Mottaghi, Yun Liu, Eric Topol, Jeff Dean, and Richard Socher. Deep learning-enabled medical com- puter vision. NPJ digital medicine, 4(1):5, 2021. 1

    2021

  4. [12]

    Pranet: Parallel reverse attention network for polyp segmentation

    Deng-Ping Fan, Ge-Peng Ji, Tao Zhou, Geng Chen, Huazhu Fu, Jianbing Shen, and Ling Shao. Pranet: Parallel reverse attention network for polyp segmentation. In International conference on medical image computing and computer- assisted intervention, pages 263–273. Springer, 2020. 2

    2020

  5. [13]

    Joint optic disc and cup seg- mentation based on multi-label deep network and polar trans- formation

    Huazhu Fu, Jun Cheng, Yanwu Xu, Damon Wing Kee Wong, Jiang Liu, and Xiaochun Cao. Joint optic disc and cup seg- mentation based on multi-label deep network and polar trans- formation. IEEE transactions on medical imaging , 37(7): 1597–1605, 2018. 2

    2018

  6. [14]

    Robust compressed sensing mri with deep generative priors

    Ajil Jalal, Marius Arvinte, Giannis Daras, Eric Price, Alexandros G Dimakis, and Jon Tamir. Robust compressed sensing mri with deep generative priors. Advances in Neural Information Processing Systems, 34:14938–14954, 2021. 2

    2021

  7. [15]

    Kvasir-seg: A segmented polyp dataset

    Debesh Jha, Pia H Smedsrud, Michael A Riegler, P ˚al Halvorsen, Thomas De Lange, Dag Johansen, and H˚avard D Johansen. Kvasir-seg: A segmented polyp dataset. In In- ternational conference on multimedia modeling, pages 451–

  8. [16]

    Survey on deep learning with class imbalance

    Justin M Johnson and Taghi M Khoshgoftaar. Survey on deep learning with class imbalance. Journal of big data , 6 (1):1–54, 2019. 1

    2019

  9. [17]

    A style-based generator architecture for generative adversarial networks

    Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 4401–4410, 2019. 2

    2019

  10. [18]

    Gans for medical image analysis

    Salome Kazeminia, Christoph Baur, Arjan Kuijper, Bram Van Ginneken, Nassir Navab, Shadi Albarqouni, and Anir- ban Mukhopadhyay. Gans for medical image analysis. Arti- ficial intelligence in medicine, 109:101938, 2020. 1, 2

    2020

  11. [19]

    Diffusion models in medical imaging: A comprehensive survey

    Amirhossein Kazerouni, Ehsan Khodapanah Aghdam, Moein Heidari, Reza Azad, Mohsen Fayyaz, Ilker Haci- haliloglu, and Dorit Merhof. Diffusion models in medical imaging: A comprehensive survey. Medical image analysis, 88:102846, 2023. 1, 2

    2023

  12. [20]

    Auto-encoding vari- ational bayes, 2013

    Diederik P Kingma, Max Welling, et al. Auto-encoding vari- ational bayes, 2013. 2

    2013

  13. [21]

    Stable diffusion segmentation for biomed- ical images with single-step reverse process

    Tianyu Lin, Zhiguang Chen, Zhonghao Yan, Weijiang Yu, and Fudan Zheng. Stable diffusion segmentation for biomed- ical images with single-step reverse process. InInternational Conference on Medical Image Computing and Computer- Assisted Intervention, pages 656–666. Springer, 2024. 6

    2024

  14. [22]

    A survey on deep learning in medical image analysis

    Geert Litjens, Thijs Kooi, Babak Ehteshami Bejnordi, Ar- naud Arindra Adiyoso Setio, Francesco Ciompi, Mohsen Ghafoorian, Jeroen Awm Van Der Laak, Bram Van Gin- neken, and Clara I S ´anchez. A survey on deep learning in medical image analysis. Medical image analysis, 42:60–88,

  15. [23]

    A multimodal comparison of latent denois- ing diffusion probabilistic models and generative adversarial networks for medical image synthesis

    Gustav M ¨uller-Franzes, Jan Moritz Niehues, Firas Khader, Soroosh Tayebi Arasteh, Christoph Haarburger, Christiane Kuhl, Tianci Wang, Tianyu Han, Teresa Nolte, Sven Nebelung, et al. A multimodal comparison of latent denois- ing diffusion probabilistic models and generative ad...

    2023

  16. [24]

    Refuge challenge: A unified framework for evaluat- ing automated methods for glaucoma assessment from fun- dus photographs

    Jos ´e Ignacio Orlando, Huazhu Fu, Jo ˜ao Barbosa Breda, Karel Van Keer, Deepti R Bathula, Andr ´es Diaz-Pinto, Ruogu Fang, Pheng-Ann Heng, Jeyoung Kim, JoonHo Lee, et al. Refuge challenge: A unified framework for evaluat- ing automated methods for glaucoma assessment from fun...

    2020

  17. [25]

    Deep structural causal models for tractable counterfactual in- ference

    Nick Pawlowski, Daniel Coelho de Castro, and Ben Glocker. Deep structural causal models for tractable counterfactual in- ference. Advances in neural information processing systems, 33:857–869, 2020. 2

    2020

  18. [26]

    Sdxl: Improving latent diffusion mod- els for high-resolution image synthesis

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas M ¨uller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion mod- els for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 5

    2023 arXiv

  19. [27]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. 1, 2

    2022

  20. [28]

    U- net: Convolutional networks for biomedical image segmen- tation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U- net: Convolutional networks for biomedical image segmen- tation. In Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, pa...

    2015

  21. [29]

    Waldstein, Ursula Schmidt-Erfurth, and Georg Langs

    Thomas Schlegl, Philipp Seeb ¨ock, Sebastian M. Waldstein, Ursula Schmidt-Erfurth, and Georg Langs. Unsupervised anomaly detection with generative adversarial networks to guide marker discovery. In Information Processing in Medi- cal Imaging, pages 146–157, Cham, 2017. Springe...

    2017

  22. [30]

    Natural synthetic anomalies for self-supervised anomaly detection and localization

    Hannah M Schl ¨uter, Jeremy Tan, Benjamin Hou, and Bern- hard Kainz. Natural synthetic anomalies for self-supervised anomaly detection and localization. In European Conference on Computer Vision, pages 474–489. Springer, 2022. 2

    2022

  23. [31]

    Mad- clip: Few-shot medical anomaly detection with clip

    Mahshid Shiri, Cigdem Beyan, and Vittorio Murino. Mad- clip: Few-shot medical anomaly detection with clip. arXiv preprint arXiv:2506.23810, 2025. 2

    2025 arXiv

  24. [32]

    A survey on image data augmentation for deep learning

    Connor Shorten and Taghi M Khoshgoftaar. A survey on image data augmentation for deep learning. Journal of big data, 6(1):1–48, 2019. 1, 2

    2019

  25. [33]

    Toward embedded detection of polyps in wce images for early diagnosis of colorectal can- cer

    Juan Silva, Aymeric Histace, Olivier Romain, Xavier Dray, and Bertrand Granado. Toward embedded detection of polyps in wce images for early diagnosis of colorectal can- cer. International journal of computer assisted radiology and surgery, 9:283–293, 2014. 2

    2014

  26. [34]

    Stepwise feature fusion: Local guides global

    Jinfeng Wang, Qiming Huang, Feilong Tang, Jia Meng, Jion- glong Su, and Sifan Song. Stepwise feature fusion: Local guides global. In International conference on medical image computing and computer-assisted intervention , pages 110–

  27. [35]

    Preparing medical imaging data for machine learning

    Martin J Willemink, Wojciech A Koszek, Cailin Hardell, Jie Wu, Dominik Fleischmann, Hugh Harvey, Les R Folio, Ronald M Summers, Daniel L Rubin, and Matthew P Lun- gren. Preparing medical imaging data for machine learning. Radiology, 295(1):4–15, 2020. 1

    2020

  28. [36]

    Diffusion models for implicit image segmentation ensembles

    Julia Wolleb, Robin Sandk ¨uhler, Florentin Bieder, Philippe Valmaggia, and Philippe C Cattin. Diffusion models for implicit image segmentation ensembles. In International Conference on Medical Imaging with Deep Learning, pages 1336–1348. PMLR, 2022. 1, 2

    2022

  29. [37]

    Medsegdiff: Medical image segmentation with diffusion probabilistic model

    Junde Wu, Rao Fu, Huihui Fang, Yu Zhang, Yehui Yang, Haoyi Xiong, Huiying Liu, and Yanwu Xu. Medsegdiff: Medical image segmentation with diffusion probabilistic model. In Medical Imaging with Deep Learning , pages 1623–1639. PMLR, 2024. 2, 6

    2024

  30. [38]

    Medsegdiff-v2: Diffusion-based medical im- age segmentation with transformer

    Junde Wu, Wei Ji, Huazhu Fu, Min Xu, Yueming Jin, and Yanwu Xu. Medsegdiff-v2: Diffusion-based medical im- age segmentation with transformer. In Proceedings of the AAAI conference on artificial intelligence, pages 6030–6038,

  31. [39]

    Generative adversar- ial network in medical imaging: A review

    Xin Yi, Ekta Walia, and Paul Babyn. Generative adversar- ial network in medical imaging: A review. Medical image analysis, 58:101552, 2019. 1, 2

    2019

  32. [40]

    Medical visual question answering via conditional rea- soning

    Li-Ming Zhan, Bo Liu, Lu Fan, Jiaxin Chen, and Xiao-Ming Wu. Medical visual question answering via conditional rea- soning. In Proceedings of the 28th ACM International Con- ference on Multimedia, pages 2345–2354, 2020. 2

    2020

  33. [41]

    Large-scale domain-specific pre- training for biomedical vision-language processing

    Sheng Zhang, Yanbo Xu, Naoto Usuyama, Jaspreet Bagga, Robert Tinn, Sam Preston, Rajesh Rao, Mu Wei, Naveen Valluri, Cliff Wong, et al. Large-scale domain-specific pre- training for biomedical vision-language processing. arXiv preprint arXiv:2303.00915, 2(3):6, 2023. 2

    2023 arXiv

  34. [462]

    Springer, 2019. 5, 6

    2019

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.