Pith. sign in

REVIEW 4 major objections 5 minor 51 references

A dual-output diffusion transformer generates OCT images and their masks together, beating DDPM and LDM baselines on fidelity and downstream segmentation.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 09:03 UTC pith:VL6VMV2G

load-bearing objection A useful, clearly written application of DiTs to joint OCT/mask synthesis; the contribution is real but the quantitative evidence is thinner than the abstract implies, mainly due to missing error bars and unvalidated mask-latent fidelity. the 4 major comments →

arxiv 2607.29337 v1 pith:VL6VMV2G submitted 2026-07-31 cs.CV cs.AI

DualDiT: A Conditional Dual-Output Diffusion Transformer for Joint OCT Image and Segmentation Mask Generation

Fernando Garc\'ia-Torres , Roc\'io del Amor , Sandra Morales , \'Alvaro Barroso , Peter Heiduschka , Bj\"orn Kemper , Valery Naranjo This is my paper
classification cs.CV cs.AI
keywords dual-output diffusion transformerjoint image-mask generationoptical coherence tomographyretinal layer segmentationsynthetic data augmentationex vivo mouse retinalatent diffusionFréchet Inception Distance
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper sets out to show that a diffusion transformer can learn the joint distribution of OCT images and their segmentation masks in a domain where annotated data are scarce and manual delineation is costly: ex vivo mouse retinal OCT. DualDiT encodes image and mask into a shared latent space via a frozen pretrained VAE, concatenates the two latents, and runs a single conditional diffusion trajectory over the joint tensor; the clean joint latent is then split and decoded into an image-mask pair. The authors report that DualDiT beats a pixel-space DDPM and a latent LDM on FID (56.14 vs 164.55 and 102.21) and sFID, and that it is the only method whose synthetic pairs consistently improve a downstream U-Net segmentation baseline, raising global Dice from 0.908 to 0.927 at +1200 synthetic samples. A three-expert panel misclassified 46% of synthetic samples as real. If these claims hold, transformer-based joint diffusion is a viable data-augmentation route for annotation-scarce medical imaging.

Core claim

The central claim is that joint image-mask generation benefits from moving the denoiser from a convolutional U-Net to a diffusion transformer while keeping both modalities in one shared latent and one denoising trajectory. DualDiT concatenates the VAE latents of image and mask channel-wise, conditions the DiT backbone on the embedding-medium class via adaLN-Zero, and denoises the 8-channel joint latent; after the trajectory, the latent is split into image and mask components and each is independently decoded. The paper reports that this design achieves the best generative fidelity among the compared diffusion baselines and, more importantly, that its synthetic pairs are the only ones that co

What carries the argument

The central mechanism is the joint latent tensor z0 = [zx ‖ zm], formed by channel-wise concatenation of the frozen VAE latents of the OCT image and its binary mask. A single DiT denoiser — transformer blocks with adaLN-Zero conditioning on timestep and class embeddings — predicts the noise on this 8-channel tensor, so the image and mask share one diffusion trajectory and are forced to be denoised coherently. At inference the recovered joint latent is split and each half is decoded by the same frozen VAE decoder. Because both outputs come from one trajectory, the paper argues, spatial and anatomical coincidence of image and mask is preserved by construction.

Load-bearing premise

The whole joint-generation claim rests on the assumption that a VAE trained on natural images provides a faithful and balanced latent representation for both greyscale OCT B-scans and binary masks, so that the shared denoising trajectory actually preserves anatomical alignment once decoded.

What would settle it

Reconstruct real masks with the frozen VAE encoder-decoder and measure Dice between input mask and reconstruction; if thin URCL boundaries are consistently lost or blurred in latent space (e.g., Dice well below the 0.9+ range achieved downstream), then the joint latent trajectory cannot be the source of the reported anatomical alignment, and the augmentation gains would not transfer to a VAE that faithfully encodes masks.

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

If this is right

  • DualDiT's synthetic OCT image-mask pairs can serve as data augmentation: at +1200 pairs, held-out segmentation Dice rises from 0.908 to 0.927 and IoU from 0.838 to 0.868, with gains concentrated on the harder resin-embedded subset.
  • Transformer-based joint diffusion outperforms the tested U-Net DDPM and LDM baselines on both FID and sFID, globally and per embedding medium, so the DiT backbone is a competitive architecture for structured image-plus-label synthesis.
  • Conditioning on embedding medium is sufficient for the model to reproduce domain-specific appearance: experts classified resin vs physiological media on synthetic images at rates statistically indistinguishable from real ones (p > 0.5).
  • The approach is scalable to higher resolutions at increased compute cost, so the same architecture extends beyond the 512×256 setting studied here.

Where Pith is reading between the lines

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

  • Since the paper does not evaluate how well the frozen VAE reconstructs binary masks, a natural extension is to measure mask latent fidelity; if the VAE distorts thin URCL boundaries, the joint-trajectory advantage may be partly an artifact of the latent space rather than the DiT itself.
  • The channel-concatenation design invites a direct ablation: training the same DiT with separate image and mask denoisers that share no joint trajectory would quantify how much of the alignment gain actually comes from the joint latent versus the transformer backbone.
  • The reported gains are measured against a strong baseline (Dice 0.908); on datasets with weaker baselines or more diverse anatomies, the augmentation benefit could plausibly be larger, which is testable on accessible human OCT datasets.
  • The expert misclassification rate of ~46% for synthetic samples came from a 100-pair sample; a larger blinded study would test whether perceptual realism persists at scale.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes DualDiT, a diffusion-transformer model that jointly generates OCT B-scans of ex vivo mouse retina and binary segmentation masks of the upper retinal cell layers, conditioned on the embedding medium (resin vs. physiological). Both image and mask are encoded with the frozen sd-vae-ft-ema VAE, concatenated channel-wise, denoised as a single 8-channel latent tensor, then split and independently decoded. The authors compare DualDiT against adapted DDPM and LDM baselines, reporting lower FID/sFID, a downstream U-Net segmentation improvement when synthetic pairs are added to the training set, and an expert evaluation in which 46% of synthetic pairs were mistaken for real. The main claims are that transformer-based joint diffusion beats U-Net-based DDPM/LDM baselines and that the synthetic pairs provide a consistent augmentation benefit.

Significance. If the claims hold, the paper makes a useful contribution by extending DiT-based joint image-mask generation to a scarce-data medical imaging setting and by providing a systematic augmentation study with a held-out volume-level split, code release, and expert assessment. The per-B-scan analysis and the sweep over augmentation sizes are strengths. However, the central empirical claims currently rest on a small dataset, single-run comparisons, and an unvalidated assumption that a natural-image VAE latent space faithfully represents binary masks and preserves image-mask alignment. These issues are addressable and do not invalidate the approach, but they must be fixed before the conclusions can be considered established.

major comments (4)
  1. [§3, Eq. (5); §4.2] The joint-generation claim rests on two unvalidated premises: that the frozen sd-vae-ft-ema codec trained on natural images reconstructs binary masks with high fidelity, and that channel-wise concatenation (Eq. 1) followed by independent decoding (Eq. 5) preserves spatial alignment between generated image and mask. No mask latent reconstruction metric is reported. Table 3 scores images only; Table 4 is indirect and the confidence-weighted teacher loss down-weights uncertain synthetic regions, so modest Dice gains can coexist with systematically misaligned masks. Please add (i) encode-decode Dice/IoU of real masks, (ii) a quantitative alignment measure between generated image-mask pairs (e.g., distance from mask boundary to predicted URCL edges), and (iii) failure cases.
  2. [§5.2, Table 4] The central augmentation claim ("consistently improves the Baseline") is not statistically supported. Best DualDiT global Dice 0.927±0.048 vs. Baseline 0.908±0.066 and IoU 0.868±0.081 vs. 0.838±0.106; the intervals overlap. Table 4 reports one training run per condition (no seeds), and no paired test on the 61 test B-scans is given even though Fig. A.7 provides per-B-scan data. Report paired significance tests (e.g., Wilcoxon signed-rank), CIs on the mean difference, and re-runs with different seeds; otherwise the 2-3.5% improvement could be noise.
  3. [§4.2, Training hyper-parameters] The three compared models use different total epochs (DDPM 6000, LDM 9500, DualDiT 5000), described as achieving "optimal hyper-parameters" with no selection criterion given. If epoch/checkpoint selection used the same FID or downstream metric employed for the final comparison, the comparison is biased. State the model-selection protocol (e.g., validation FID, fixed checkpoint schedule) and ideally match compute or show a learning curve for baselines. Without this, the FID advantage in Table 3 cannot be attributed to architecture.
  4. [§5.3, Figure 5] The expert study's headline percentages aggregate only 50 synthetic and 50 real pairs judged by three experts, and no chance-level comparison or confidence intervals are given for the 46%/42% misclassification rates. Reporting per-expert counts and a binomial test would clarify whether the rates differ from chance. The embedding-medium AUC comparisons (Table 5) have wide uncertainty despite non-significant p-values.
minor comments (5)
  1. [Table 3] FID/sFID are point estimates with no seeds or confidence intervals; state the number of generated samples used and repeat sampling seeds.
  2. [§5.2 and §6] The relative improvement is given as 2.09%/3.58% in §5.2 and 2.07%/3.50% in §6; harmonize the numbers.
  3. [Eq. (7)] The notation for the BCE term is undefined; clarify the exact form of L_BCE and how α=0.5 weights it against Dice.
  4. [§4.3.1] sFID is not precisely defined; cite the exact implementation and state whether it uses the same Inception features as FID.
  5. [Figures A.1-A.6] The qualitative overlays are not an alignment check; add scale bars and, if possible, quantitative edge-error measurements.

Circularity Check

0 steps flagged

No significant circularity: DualDiT's claims are empirical and evaluated on held-out real data; remaining concerns are validation gaps, not reductions to inputs.

full rationale

Walking the paper's derivation chain, no load-bearing step reduces to its own inputs. The joint latent construction z0=[zx||zm] (Eq. 1) and the split-and-decode step (Eq. 5) are design choices, not derived predictions; the claim that they preserve anatomical coherence is asserted qualitatively, and the unmeasured fidelity of the frozen natural-image VAE for binary masks is a validity risk, not a circular reduction. Training optimizes the standard noise-prediction objective (Eq. 3) with no fitted parameter later renamed as a prediction. Generative quality is assessed with externally defined FID/sFID using a pretrained Inception network and OpenAI's evaluation pipeline, not with quantities derived from the model's own training objective. Downstream segmentation is evaluated on a strictly held-out real test set, with all augmentation methods (DDPM, LDM, DualDiT) treated under the same training protocol and the same teacher-student confidence weighting. The self-citations present in the paper ([19], [20], [23], [24], [42], [49]) support data provenance, background, and a method borrowed from a different application; none is invoked as a uniqueness theorem, none forbids alternatives, and none supplies the central generative or segmentation claim. The paper's own limitation statements acknowledge small dataset size and the need for further validation, which reinforces that the reported improvements are empirical rather than forced by construction. No circular step was found.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The central claim depends on standard diffusion machinery, a pretrained VAE used out-of-domain, and dataset assumptions. There are no fitted physical parameters or new entities, but several hand-chosen hyperparameters and a post-hoc augmentation-size selection affect the headline numbers.

free parameters (5)
  • n_synth (augmentation set size) = 1200 (best global Dice/IoU)
    Chosen from a sweep of 50–1400 synthetic pairs; the headline improvement corresponds to this selected value, not a pre-registered protocol.
  • Noise schedule endpoints = beta_start=1.5e-4, beta_end=1.95e-2 (linear)
    Hand-chosen standard schedule; shared across models but not swept, and it affects sample quality.
  • CFG unconditional probability = 0.1
    Hand-set for classifier-free guidance; influences the fidelity/diversity trade-off.
  • Training epochs per model = DualDiT 5000, DDPM 6000, LDM 9500
    Each baseline was trained until its 'optimal hyper-parameter combination'; the comparison is not compute- or epoch-matched.
  • EMA decay = 0.9999
    Hand-chosen exponential moving average decay used for sampling; affects output quality.
axioms (5)
  • standard math The DDPM forward/reverse process and noise-prediction objective (Eq. 3) with T=1000 produce valid samples.
    Adopted from Ho et al. [13]; standard diffusion theory.
  • domain assumption A VAE trained on natural images (sd-vae-ft-ema) is a valid encoder/decoder for OCT images and binary masks.
    Section 3/4.2: the VAE is frozen and used out-of-domain; no reconstruction evaluation for masks is provided.
  • domain assumption Channel-wise concatenation of image and mask latents in one denoising trajectory is sufficient to enforce anatomical alignment.
    This is the core design choice (Eq. 1); no ablation against separate streams or cross-attention is given.
  • domain assumption Inception-based FID/sFID is a meaningful realism metric for greyscale OCT images.
    Section 4.3.1 uses the standard ImageNet-pretrained Inception pipeline; not validated for this imaging domain.
  • domain assumption Volume-level splitting prevents inter-slice leakage, and manual URCL masks are correct ground truth.
    Section 4.1 and Table 2; reasonable experimental hygiene, but depends on annotation quality and limited volume counts.

pith-pipeline@v1.3.0-daily-deepseek · 21397 in / 11860 out tokens · 131672 ms · 2026-08-03T09:03:12.963558+00:00 · methodology

0 comments
read the original abstract

Background and Objective: Generating realistic medical images with anatomically accurate segmentation masks helps address the shortage of annotated data in medical imaging, particularly in optical coherence tomography (OCT) of mouse eyes, where manual retinal layer delineation is labour-intensive due to tiny structures and required expertise, resulting in scarce datasets. While diffusion models perform well in medical image synthesis, joint image-mask generation has relied mainly on U-Net-based denoisers, leaving diffusion transformers largely unexplored. Methods: We propose a conditional dual-output Diffusion Transformer (DualDiT) for joint synthesis of OCT B-scans and segmentation masks of the upper retinal cell layers in ex vivo mouse retina. DualDiT encodes both modalities into a shared latent space via a pretrained VAE, concatenates their latent representations, and performs conditional diffusion over the joint tensor. We compared DualDiT against two adapted diffusion baselines: DDPM and LDM. Generative quality was assessed via Fr\'echet Inception Distance (FID) and spatial FID (sFID); practical utility via synthetic data augmentation for downstream U-Net segmentation; and perceptual realism via evaluation by three domain experts. Results: DualDiT achieved the best generative quality (FID 56.14, sFID 114.35), outperforming DDPM and LDM. Expert panels misclassified 46% of synthetic samples as real and 42% of real samples as synthetic. Adding DualDiT-generated images and masks improved Dice and IoU scores on a held-out segmentation test set. Conclusions: DualDiT shows that transformer-based diffusion models can effectively learn the joint distribution of OCT images and segmentation masks, surpassing DDPM- and LDM-based baselines in generative fidelity, downstream utility, and perceptual realism, highlighting its potential for data augmentation in annotation-scarce medical imaging.

Figures

Figures reproduced from arXiv: 2607.29337 by \'Alvaro Barroso, Bj\"orn Kemper, Fernando Garc\'ia-Torres, Peter Heiduschka, Roc\'io del Amor, Sandra Morales, Valery Naranjo.

Figure 1
Figure 1. Figure 1: Overview of the DualDiT framework. Paired image and segmentation mask inputs (𝑥, 𝑚) are encoded (𝑓𝐸 (.)) into a shared latent space and concatenated along the channel dimension to form a joint latent tensor 𝑧0 = [𝑧𝑥‖𝑧𝑚 ]. During training, Gaussian noise (𝜖) is progressively added to 𝑧0 over 𝑇 timesteps following the forward process 𝑞(𝑧𝑡 ∣ 𝑧0 , 𝑡). At each timestep 𝑡, the noisy latent 𝑧𝑡 , the timestep embe… view at source ↗
Figure 2
Figure 2. Figure 2: Examples of URCL layer segmentation for representative B-scans of fluid and resin embedded samples. The images show the upper and lower boundaries of the URCL layer across the two experimental groups (see [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Example figure of a synthetic resin-embedded sample for expert evaluation: (Left) OCT B-Scan; (Middle) URCL segmentation mask; (Right) URCL delineation on the B-Scan. 100 images (50 real and 50 synthetic) were analysed. The two embedding classes were equally represented, yielding 25 images per class and per sample type (real or synthetic). To ensure the samples were assessed correctly, in each case, the B-… view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of real and synthetic images with their corresponding URCL mask overlays. The left two columns show physiologically embedded samples, whereas the right two columns show resin-embedded samples. relative to the Baseline. The model trained with DualDiT￾generated samples achieves a global Dice score of 0.917 ± 0.058 and an IoU of 0.852 ± 0.095, compared with 0.908 ± 0.066 and 0.838 ± 0.106 for the B… view at source ↗
Figure 5
Figure 5. Figure 5: Confusion matrix for the classification performed by experts. (Left) Classification for real and synthetic samples; (Middle) Embedding medium classification for real samples; (Right) Embedding medium classification for synthetic pairs. The results highlight that DualDiT outperforms the Baseline in the majority of cases, although performance decreases are observed for a small subset of B-scans. 5.3. Externa… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

51 extracted references · 14 canonical work pages

  1. [1]

    Kebaili, J

    A. Kebaili, J. Lapuyade-Lahorgue, S. Ruan, Deep learning ap- proachesfordataaugmentationinmedicalimaging:Areview,Journal of Imaging 9 (2023). doi:10.3390/jimaging9040081

    work page doi:10.3390/jimaging9040081 2023

  2. [2]

    Borrajo, A survey on deep learning applied to medical images: fromsimpleartificialneuralnetworkstogenerativemodels35(2023) 2291–2323

    P.Celard,E.L.Iglesias,J.M.Sorribes-Fdez,R.Romero,A.S.Vieira, L. Borrajo, A survey on deep learning applied to medical images: fromsimpleartificialneuralnetworkstogenerativemodels35(2023) 2291–2323. doi:10.1007/s00521-022-07953-4

    work page doi:10.1007/s00521-022-07953-4 2023

  3. [3]

    Chlap, H

    P. Chlap, H. Min, N. Vandenberg, J. Dowling, L. Holloway, A. Ha- worth, A review of medical image data augmentation techniques for deep learning applications 65 (2021) 545–563. doi:10.1111/1754-948 5.13261

    work page doi:10.1111/1754-948 2021

  4. [4]

    Islam, M

    S. Islam, M. T. Aziz, H. R. Nabil, J. R. Jim, M. F. Mridha, M. M. Kabir, N. Asai, J. Shin, Generative adversarial networks (GANs) in medical imaging: Advancements, applications, and challenges 12 (2024) 35728–35753. doi:10.1109/ACCESS.2024.3370848

    arXiv 2024

  5. [5]

    doi:10.1016/ j.ijmedinf.2025.105903

    C.Oulmalme,H.Nakouri,F.Jaafar,Asystematicreviewofgenerative AI approaches for medical image enhancement: Comparing GANs, transformers, and diffusion models 199 (2025) 105903. doi:10.1016/ j.ijmedinf.2025.105903

    arXiv 2025

  6. [6]

    Waheed, M

    A. Waheed, M. Goyal, D. Gupta, A. Khanna, F. Al-Turjman, P. R. Pinheiro, Covidgan: data augmentation using auxiliary classifier gan forimprovedcovid-19detection, IeeeAccess8(2020)91916–91923. doi:10.1109/ACCESS.2020.2994762

    arXiv 2020

  7. [7]

    Isola, J.-Y

    P. Isola, J.-Y. Zhu, T. Zhou, A. A. Efros, Image-to-image translation with conditional adversarial networks, 2017, pp. 1125–1134. URL: https://openaccess.thecvf.com/content_cvpr_2017/html/Isola_Image -To-Image_Translation_With_CVPR_2017_paper.html

    2017

  8. [8]

    G. Yang, S. Yu, H. Dong, G. Slabaugh, P. L. Dragotti, X. Ye, F. Liu, S. Arridge, J. Keegan, Y. Guo, et al., Dagan: deep de- aliasing generative adversarial networks for fast compressed sensing mri reconstruction, IEEE transactions on medical imaging 37 (2017) 1310–1321. doi:https://doi.org/10.1109/TMI.2017.2785879

    arXiv 2017

  9. [9]

    D. P. Kingma, M. Welling, An introduction to variational autoen- coders 12 (2019) 307–392. doi:10.1561/2200000056

    doi:10.1561/2200000056 2019

  10. [10]

    Rguibi, A

    Z. Rguibi, A. Hajami, D. Zitouni, M. Yassine, A. Elqaraoui, Medical variational autoencoder and generative adversarial network for medi- cal imaging 32 (2023) 1–1x. doi:10.11591/ijeecs.v32.i1.pp494-505

    work page doi:10.11591/ijeecs.v32.i1.pp494-505 2023

  11. [11]

    Y. Shi, A. Abulizi, H. Wang, K. Feng, N. Abudukelimu, Y. Su, H. Abudukelimu, Diffusion models for medical image computing: A survey 30 (2025) 357–383. doi:10.26599/TST.2024.9010047

    arXiv 2025

  12. [12]

    Müller-Franzes, J

    G. Müller-Franzes, J. M. Niehues, F. Khader, S. T. Arasteh, C. Haar- burger, C. Kuhl, T. Wang, T. Han, T. Nolte, S. Nebelung, et al., A multimodal comparison of latent denoising diffusion probabilistic models and generative adversarial networks for medical image syn- thesis, Scientific reports 13 (2023) 12098. doi:https://doi.org/10.1 038/s41598-023-39278-0

    2023

  13. [13]

    J. Ho, A. Jain, P. Abbeel, Denoising diffusion probabilistic models, in: Advances in Neural Information Processing Systems, volume 33, Curran Associates, Inc., 2020, pp. 6840–6851. URL:https://procee dings.neurips.cc/paper_files/paper/2020/file/4c5bcfec8584af0d967 f1ab10179ca4b-Paper.pdf

    2020

  14. [14]

    Rombach, A

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, B. Ommer, High- resolution image synthesis with latent diffusion models, 2022. doi:10 .48550/arXiv.2112.10752.arXiv:2112.10752 [cs]

    Pith/arXiv arXiv doi:10.48550/arxiv.2112.10752 2022

  15. [15]

    W. Wang, J. Xia, G. Luo, S. Dong, X. Li, J. Wen, S. Li, Diffusion model for medical image denoising, reconstruction and translation 124 (2025) 102593. doi:10.1016/j.compmedimag.2025.102593

    arXiv 2025

  16. [16]

    Chataut, S

    S. Chataut, S. Bhatta, B. Dahal, G. Ojha, S. Raut, B. Subedi, B. Bas- takoti, Generative artificial intelligence in healthcare: A system- atic review of GANs, diffusion models, large language models, and variational autoencoders for medical applications 3 (2025) 182–203. doi:10.59324/ejaset.2025.3(4).16

    work page doi:10.59324/ejaset.2025.3(4).16 2025

  17. [17]

    A. F. Fercher, W. Drexler, C. K. Hitzenberger, T. Lasser, Optical co- herencetomography-principlesandapplications, Reportsonprogress in physics 66 (2003) 239–303. doi:10.1088/0034-4885/66/2/204

    work page doi:10.1088/0034-4885/66/2/204 2003

  18. [18]

    M. A. Mayer, J. Hornegger, C. Y. Mardin, R. P. Tornow, Retinal nervefiberlayersegmentationonfd-octscansofnormalsubjectsand glaucoma patients, Biomedical optics express 1 (2010) 1358–1383. doi:https://doi.org/10.1364/BOE.1.001358

    work page doi:10.1364/boe.1.001358 2010

  19. [19]

    Morales, A

    S. Morales, A. Colomer, J. M. Mossi, R. del Amor, D. Woldbye, K. Klemp, M. Larsen, V. Naranjo, Retinal layer segmentation in rodent OCT images: Local intensity profiles & fully convolutional neuralnetworks198(2021)105788.doi:10.1016/j.cmpb.2020.105788

    arXiv 2021

  20. [20]

    R. D. Amor, S. Morales, A. N. Colomer, J. M. Mossi, D. Woldbye, K. Klemp, M. Larsen, V. Naranjo, Towards automatic glaucoma assessment: An encoder-decoder CNN for retinal layer segmentation in rodent OCT images, in: 2019 27th European Signal Processing Conference (EUSIPCO), IEEE, 2019, pp. 1–5. doi:10.23919/EUSIPCO .2019.8902794

    arXiv 2019

  21. [21]

    R. S. Allen, K. Bales, A. Feola, M. T. Pardue, In vivo structural assessments of ocular disease in rodent models using optical coher- ence tomography, Journal of visualized experiments: JoVE (2020) 10–3791. doi:10.3791/61588

    work page doi:10.3791/61588 2020

  22. [22]

    Tschernig, L

    T. Tschernig, L. Thrane, T. M. Jørgensen, J. Thommes, R. Pabst, T. Yelbuz, An elegant technique for ex vivo imaging in experimental research—optical coherence tomography (oct), Annals of Anatomy- Anatomischer Anzeiger 195 (2013) 25–27. doi:https://doi.org/10.1 016/j.aanat.2012.07.005

    2013

  23. [23]

    Barroso, P

    A. Barroso, P. Heiduschka, G. Nettels-Hackert, S. Ketelhut, R. d. Amor,F.García-Torres,S.Morales-Martínez,V.Naranjo,B.Kemper, J.Schnekenburger, Durableexvivomouseretina3dtissuemodelsfor optical coherence tomography, in: Label-free Biomedical Imaging and Sensing (LBIS) 2024, volume 12854, SPIE, 2024, pp. 21–23. doi:10.1117/12.3002538

    work page doi:10.1117/12.3002538 2024

  24. [24]

    García-Torres, R

    F. García-Torres, R. del Amor, S. Morales-Martínez, A. Barroso, B. Kemper, J. Schnekenburger, V. Naranjo, Using diffusion models for data augmentation on limited rodent oct datasets, in: Intelligent Data Engineering and Automated Learning – IDEAL 2024: 25th International Conference, Valencia, Spain, November 20–22, 2024, Proceedings, Part I, Springer-Verl...

    work page doi:10.1007/978-3-031-77731-8_29 2024

  25. [25]

    Y. Tian, A. McCarthy, G. Gomide, N. Liddle, J. Golebka, R. Chen, J. Liebmann, K. Thakoor, OCTDiff: Bridged diffusion model for portable OCT super-resolution and enhancement 38 (2026) 41445– 41465

    2026

  26. [26]

    doi:10.1109/TBME.2025.3556794

    N.Abbasi,A.Wong,K.Bizheva,Aphysics-informeddiffusionmodel for super-resolved reconstruction of optical coherence tomography data 72 (2025) 2937–2946. doi:10.1109/TBME.2025.3556794

    arXiv 2025

  27. [27]

    doi:10.3390/jimaging10040086

    H.Ahmed,Q.Zhang,R.Donnan,A.Alomainy, Denoisingofoptical coherencetomographyimagesinophthalmologyusingdeeplearning: A systematic review 10 (2024) 86. doi:10.3390/jimaging10040086. F. García-Torres et al.:Preprint submitted to ElsevierPage 13 of 21 Conditional Dual-Output Diffusion Transformer (DualDiT)

    work page doi:10.3390/jimaging10040086 2024

  28. [28]

    S. Li, M. Dan, Y. Chu, J. Yu, Y. Zhao, P. Zhao, RetiDiff: Diffusion- based synthesis of retinal OCT images for enhanced segmentation, in: J. C. Gee, D. C. Alexander, J. Hong, J. E. Iglesias, C. H. Sudre, A. Venkataraman, P. Golland, J. H. Kim, J. Park (Eds.), Medical Image Computing and Computer Assisted Intervention – MICCAI 2025, Springer Nature Switzer...

    2025

  29. [29]

    Y. Wu, W. He, D. Eschweiler, N. Dou, Z. Fan, S. Mi, P. Walter, J. Stegmaier, Retinal OCT synthesis with denoising diffusion prob- abilistic models for layer segmentation, in: 2024 IEEE International Symposium on Biomedical Imaging (ISBI), 2024, pp. 1–5. doi:10.1 109/ISBI56570.2024.10635836

    arXiv 2024

  30. [30]

    K. Du, U. Doshi, B. DiCenzo, J. Jiang, E. Wu, A. Gadari, S. C. Vupparaboina,E.Sadeghi,S.C.Bollepalli,J.-A.Sahel,J.Chhablani, K.K.Vupparaboina,Benchmarkingdiffusionmodelsagainststate-of- the-artarchitecturesforOCTfluidbiomarkersegmentation20(2025) e0335615. doi:10.1371/journal.pone.0335615

    work page doi:10.1371/journal.pone.0335615 2025

  31. [31]

    Huang, X

    K. Huang, X. Ma, Y. Zhang, N. Su, S. Yuan, Y. Liu, Q. Chen, H. Fu, Memory-efficient high-resolution OCT volume synthesis withcascadedamortizedlatentdiffusionmodels, in:M.G.Linguraru, Q. Dou, A. Feragen, S. Giannarou, B. Glocker, K. Lekadir, J. A. Schnabel (Eds.), Medical Image Computing and Computer Assisted Intervention–MICCAI2024,SpringerNatureSwitzerla...

    work page doi:10.1007/978-3-031-72104-5_46 2024

  32. [32]

    R. H. Badhon, A. C. Thompson, J. I. Lim, T. Leng, M. N. Alam, Diffusion model based OCT to OCTA translation 12 (2025). doi:10 .3389/fmed.2025.1655453

    arXiv 2025

  33. [33]

    B. Yang, J. Li, J. Wang, R. Li, K. Gu, B. Liu, DiffusionDCI: A novel diffusion-based unified framework for dynamic full-field OCT image generationandsegmentation12(2024)37702–37714.doi:10.1109/AC CESS.2024.3372863

    arXiv 2024

  34. [34]

    Toker, M

    A. Toker, M. Eisenberger, D. Cremers, L. Leal-Taixé, SatSynth: Augmenting image-mask pairs through diffusion models for aerial semantic segmentation, 2024, pp. 27695–27705. URL:https://op enaccess.thecvf.com/content/CVPR2024/html/Toker_SatSynth_Augment ing_Image-Mask_Pairs_through_Diffusion_Models_for_Aerial_Semant ic_CVPR_2024_paper.html

    2024

  35. [35]

    J. Mao, Y. Wang, Y. Tang, D. Xu, K. Wang, Y. Yang, Z. Zhou, Y. Zhou, MedSegFactory: Text-guided generation of medical image- maskpairs, 2025,pp.21525–21535.URL:https://openaccess.thecv f.com/content/ICCV2025/html/Mao_MedSegFactory_Text-Guided_Genera tion_of_Medical_Image-Mask_Pairs_ICCV_2025_paper.html

    2025

  36. [36]

    3762–3771

    Y.Frisch,C.Bornberg,M.Fuchs,A.Mukhopadhyay, GAUDA:Gen- erativeadaptiveuncertainty-guideddiffusion-basedaugmentationfor surgical segmentation, in: 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2025, pp. 3762–3771. doi:10.1109/WACV61041.2025.00370, ISSN: 2642-9381

    arXiv 2025

  37. [37]

    W. Wu, Y. Zhao, M. Z. Shou, H. Zhou, C. Shen, DiffuMask: Synthe- sizingimageswithpixel-levelannotationsforsemanticsegmentation using diffusion models, 2023, pp. 1206–1217. URL:https://openac cess.thecvf.com/content/ICCV2023/html/Wu_DiffuMask_Synthesizing_ Images_with_Pixel-level_Annotations_for_Semantic_Segmentation_U sing_ICCV_2023_paper.html

    2023

  38. [38]

    Z. Li, Q. Zhou, X. Zhang, Y. Zhang, Y. Wang, W. Xie, Open- vocabulary object segmentation with diffusion models, 2023, pp. 7667–7676. URL:https://openaccess.thecvf.com/content/ICCV 2023/html/Li_Open-vocabulary_Object_Segmentation_with_Diffusion _Models_ICCV_2023_paper.html

    2023

  39. [39]

    J. H. Park, K. Jo, S. Baik, SeeDiff: Off-the-shelf seeded mask generation from diffusion models 39 (2025) 6406–6415. doi:10.160 9/aaai.v39i6.32686

    2025

  40. [40]

    URL:https://huggin gface.co/stabilityai/sd-vae-ft-ema

    stabilityai/sd-vae-ft-ema·hugging face, 2022. URL:https://huggin gface.co/stabilityai/sd-vae-ft-ema

    2022

  41. [41]

    Peebles, S

    W. Peebles, S. Xie, Scalable diffusion models with transformers, 2023, pp. 4195–4205. URL:https://openaccess.thecvf.com/cont ent/ICCV2023/html/Peebles_Scalable_Diffusion_Models_with_Transfo rmers_ICCV_2023_paper.html

    2023

  42. [42]

    Barroso, S

    A. Barroso, S. Ketelhut, G. Nettels-Hackert, P. Heiduschka, R. d. Amor,V.Naranjo,B.Kemper,J.Schnekenburger, Durable3dmurine exvivoretinaglaucomamodelsforopticalcoherencetomography14 (2023) 4421–4438. doi:10.1364/BOE.494271

    work page doi:10.1364/boe.494271 2023

  43. [43]

    D. P. Kingma, J. Ba, Adam: A method for stochastic optimization,

  44. [44]

    J. Ho, T. Salimans, Classifier-free diffusion guidance, 2022. URL: https://arxiv.org/abs/2207.12598.arXiv:2207.12598

    Pith/arXiv arXiv 2022

  45. [45]

    Heusel, H

    M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, S. Hochreiter, GANs trained by a two time-scale update rule converge to a local nash equilibrium, in: Advances in Neural Information Processing Systems, volume 30, Curran Associates, Inc., 2017. URL:https: //proceedings.neurips.cc/paper/2017/hash/8a1d694707eb0fefe65 871369074926d-Abstract.html

    2017

  46. [46]

    Dhariwal, A

    P. Dhariwal, A. Nichol, Diffusion models beat GANs on image synthesis, 2021. doi:10.48550/arXiv.2105.05233.arXiv:2105.05233 [cs]

    Pith/arXiv arXiv doi:10.48550/arxiv.2105.05233 2021

  47. [47]

    Ronneberger, P

    O. Ronneberger, P. Fischer, T. Brox, U-net: Convolutional networks forbiomedicalimagesegmentation,in:N.Navab,J.Hornegger,W.M. Wells,A.F.Frangi(Eds.),MedicalImageComputingandComputer- Assisted Intervention – MICCAI 2015, Springer International Pub- lishing, 2015, pp. 234–241. doi:10.1007/978-3-319-24574-4_28

    doi:10.1007/978-3-319-24574-4_28 2015

  48. [48]

    doi:https://doi.org/10.1016/j.media.2021.10203 5

    J.Ma,J.Chen,M.Ng,R.Huang,Y.Li,C.Li,X.Yang,A.L.Martel, Lossodysseyinmedicalimagesegmentation, Medicalimageanalysis 71 (2021) 102035. doi:https://doi.org/10.1016/j.media.2021.10203 5

    arXiv 2021

  49. [49]

    del Amor, A

    R. del Amor, A. Colomer, S. Morales, C. Pulgarín-Ospina, L. Ter- radez, J. Aneiros-Fernandez, V. Naranjo, A self-contrastive learning framework for skin cancer detection using histological images, in: 2022 IEEE International Conference on Image Processing (ICIP), 2022, pp. 2291–2295. doi:10.1109/ICIP46576.2022.9897435

    arXiv 2022

  50. [50]

    doi:https://doi.org/10.1186/s13104-022-06096-y

    D.Müller,I.Soto-Rey,F.Kramer, Towardsaguidelineforevaluation metrics in medical image segmentation, BMC research notes 15 (2022) 210. doi:https://doi.org/10.1186/s13104-022-06096-y. A. Qualitative Results of Joint Synthetic Image and Mask Generation F. García-Torres et al.:Preprint submitted to ElsevierPage 14 of 21 Conditional Dual-Output Diffusion Trans...

    doi:10.1186/s13104-022-06096-y 2022

  51. [2017]

    doi:10.48550/arXiv.141 2.6980.arXiv:1412.6980 [cs.LG]

    URL:http://arxiv.org/abs/1412.6980. doi:10.48550/arXiv.141 2.6980.arXiv:1412.6980 [cs.LG]

    Pith/arXiv arXiv doi:10.48550/arxiv.1412.6980