Pith. sign in

REVIEW 4 major objections 5 minor 34 references

Exploring the Effectiveness of Deep Features from Domain-Specific Foundation Models in Retinal Image Synthesis

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Deep features from a retinal foundation model do not make synthetic colour fundus images more realistic or clinically accurate, while a simple edge-detection loss nearly matches a natural-image perceptual loss.

desk verdict Useful empirical comparison with a practical edge-loss takeaway, but the headline claim about domain-specific deep features is overgeneralized because the RETFound loss is confounded with architecture, layer count, and unreported loss weights. read the letter →

arxiv 2506.11753 v1 pith:SF4SKGEK submitted 2025-06-13 eess.IV cs.CV

Zuzanna Skorniewska , Bartlomiej W. Papiez This is my paper
classification eess.IVcs.CV
keywords colourfundusimagingsyntheticmedicalimagesVQ-GANperceptuallossretinalimagesynthesisfoundationmodelfeaturesedgedetectionvascularmorphology
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper asks whether using deep features from a foundation model trained on retinal images improves synthetic colour fundus image generation. The authors train VQ-GAN autoencoders with four supervision losses: pixel-only, edge-based, perceptual loss from a natural-image network, and a new loss computed in the latent space of RETFound, a vision transformer pretrained on 1.6 million retinal images. Across fidelity metrics, downstream cardiovascular risk prediction, and direct comparison of 45 retinal morphology features, the RETFound loss underperforms and does not clearly beat the pixel-only baseline. Perceptual loss leads on most criteria, and a simple Meijering edge-detection loss is nearly as good at far lower cost. The paper concludes that domain-specific foundation-model features do not yield more realistic or clinically accurate synthetic retinal images.

What carries the argument

The key object is the RETFound loss: the L1 distance in the 1024-dimensional embedding space of the last transformer block of RETFound, computed between real and synthetic images and added to a VQ-GAN training scheme alongside pixel and GAN losses. The comparison machinery includes perceptual loss, which sums L2 distances over five VGG-16 activation layers, and edge loss, which computes L2 distance between Meijering-filter vesselness responses on the green channel. Validation runs through FID, MMD, and MS-SSIM for fidelity, downstream Train-Synthetic/Test-Real prediction of cardiovascular risk factors, and permutation tests over 45 AutoMorph retinal morphology features.

What would settle it

Retrain the VQ-GAN with a RETFound-based loss that, like the perceptual loss, sums distances over several transformer layers instead of only the final layer. If FID, downstream prediction errors, and feature-overlap p-values then rise to match the perceptual-loss model, the claim that domain-specific deep features do not help would be contradicted.

Watch

Extended reading notes

Core claim

The central claim is that leveraging deep features of foundational models trained on domain data does not yield samples that are more realistic or that better preserve morphological and clinical information in colour fundus images. In head-to-head comparison, the VQ-GAN supervised by RETFound loss achieved FID 11.43 versus 2.45 for perceptual loss and 5.47 for edge loss, and in downstream clinical prediction tasks it performed worse than even the pixel-only baseline. Perceptually supervised models matched the largest share of real retinal features in permutation tests, while the RETFound-supervised model deviated most on fractal dimension and vascular complexity. The authors note that this is surprising given the established use of RETFound encoder features for cardiovascular risk prediction, and they list architectural differences, such as a vision transformer versus a CNN and reliance on a single activation layer, as possible explanations.

Load-bearing premise

The comparison assumes that the two losses differ mainly in whether the features come from a retinal foundation model or a natural-image network; in fact they also differ in number of network layers, so the disappointing results may come from the implementation rather than from domain-specific features.

Editorial extensions

If this is right

  • For colour fundus synthesis, a domain-specific foundation-model perceptual loss built from a single final layer is not worth its compute relative to perceptual or edge losses.
  • Simple edge-detection supervision is a practical, weight-free alternative that nearly matches expensive deep-feature losses for vascular sharpness.
  • Downstream clinical prediction and morphological feature overlap should be part of evaluating synthetic medical images, since fidelity metrics such as MMD and MS-SSIM alone could not separate the models.
  • No tested model matched real images in downstream prediction, so synthetic fundus data still cannot replace real data for these tasks.
  • The result leaves open whether transformer-based losses or multi-layer perceptual losses could improve synthesis, because the RETFound loss used only one activation layer.

Reading between the lines

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

  • Editorial inference: the paper does not isolate whether the failure comes from domain-specific features, from the vision-transformer architecture, or from using only one activation layer; a fair test would need a multi-layer RETFound loss or a CNN retinal backbone.
  • Editorial inference: if the single-layer implementation drives the gap, the broader conclusion that domain-specific features do not help would be too hasty and should be tested directly.
  • Editorial inference: the strong edge-loss result hints that low-level vessel geometry is the bottleneck for clinical utility, more than high-level semantic features; combining edge loss with perceptual loss could test this.
  • Editorial inference: the validation design could serve as a template for other medical modalities, reporting feature-overlap and downstream-task performance rather than relying on FID alone.
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 / 5 minor

Summary. The paper investigates whether a perceptual-style loss built on deep features of RETFound, a self-supervised foundation model trained on 1.6 million colour fundus images, improves VQ-GAN-based synthesis of retinal images compared with a VGG-16 perceptual loss, an edge-detection (Meijering filter) loss, and a pixel-only baseline. All models are trained on 90,344 UK Biobank colour fundus images at 256x256 resolution and evaluated on FID/MMD/MS-SSIM, on Train-Synthetic/Test-Real prediction of age, sex, BMI, blood pressure, HbA1c and glucose using AutoMorph-extracted features, and on distributional matching of 45 retinal morphology features via permutation tests. The authors conclude in Section 5 that deep features from a domain-specific foundation model do not yield more realistic or morphologically accurate synthetic fundus images, and that a conventional edge-based loss is a cheap and competitive alternative.

Significance. If established, the negative result is a useful contribution to medical image synthesis: it challenges the implicit assumption that domain-specific foundation-model features are the right backbone for perceptual-style losses in this setting, and it identifies edge-based supervision as a practical, low-cost alternative. The study has genuine methodological strengths: the evaluation pipeline is unusually comprehensive for the area, spanning fidelity, downstream clinical prediction, and direct morphological comparison; the evaluation metrics and feature extractors (AutoMorph, RETFound, InceptionV3) are external pretrained models independent of the training losses, so circularity is minimal; and the central claim is a clearly falsifiable negative result rather than another positive demonstration on a single metric. The main weakness is that the headline claim is stated more broadly than the experiments can support: the RETFound loss and the perceptual loss differ in architecture, layer count, distance norm, resolution, and unreported loss weights, and the paper's own Discussion concedes that these factors are plausible explanations for the observed gap.

major comments (4)
  1. [§2.1, §3.3, §3.5, §5] The central negative claim that domain-specific deep features are ineffective rests on comparing VQ-GAN (RETFound) with VQ-GAN (Perceptual), but the two losses differ along several axes at once. The RETFound loss (Section 3.3) is an L1 distance on a single 1024-dimensional final-layer embedding of a ViT-Large, computed after resizing inputs to 224x224, whereas the perceptual loss (Section 3.5) is a sum of L2 distances over five VGG-16 activation layers (4th, 9th, 16th, 23rd, 30th) at the native 256x256 resolution. Feature source, architecture (ViT vs CNN), layer count (1 vs 5), distance norm (L1 vs L2), and resolution are therefore all varied simultaneously, and the weighting of the auxiliary loss terms relative to the pixel and GAN losses is never reported (Section 2.1 lists three gradient sources without coefficients). The Discussion (Section 5) itself names the architecture difference and the layer-count difference as possible reasons for the RETFound loss underperforming, which concedes that the experiment does not isolate the effect of feature source. As stated, the headline conclusion overgeneralizes from a single confounded configuration; the authors should either add ablations that match layer count and architecture (for example, a multi-layer RETFound loss or a single-layer VGG loss) or narrow the claim to the specific loss configuration tested.
  2. [§4.2, Table 2] The sentence in Section 4.2 that 'none of the synthetic models achieved error rates as low as those based on real data' is contradicted by the paper's own Table 2: for age, VQ-GAN (Perceptual) reports MAE 7.60 (0.27), which is lower than the real-data MAE of 7.72 (0.13); for sex, VQ-GAN (RETFound) reports F1 0.53 (0.01), higher than the real-data 0.52 (0.01). The blanket statement is therefore factually incorrect as written and should be replaced with a per-variable comparison or a qualified statement such as 'in most prediction tasks.'
  3. [§4.1, Tables 1–2, §4.3] FID and MMD are reported in Table 1 as single point values with no error bars, and the number of training runs or evaluation seeds is not stated anywhere. For a paper whose main contribution is a negative result, the absence of uncertainty quantification is load-bearing: the FID spread across models (2.45, 5.47, 11.43, 19.95) and the morphological 'matches' counted in Section 4.3 (3, 3, 4, and 5 of 45 features) are small-number comparisons, and without confidence intervals, multiple seeds, or a statistical test the reader cannot judge whether the ordering of the models is stable. This is especially important for the claim that the RETFound loss does not improve over the pixel-only baseline, since its FID of 11.43 is closer to real data than the pixel baseline's 19.95 but this difference may be within run-to-run noise.
  4. [§5, Table 1] The Discussion states that 'the RETFound loss consistently underperformed,' but Table 1 shows that VQ-GAN (RETFound) achieves FID 11.43 versus 19.95 for VQ-GAN (Pixel), i.e., the RETFound loss substantially improves fidelity relative to the pixel-only baseline on this metric. The abstract's claim that domain-specific deep features 'do not improve autoencoder image generation' is likewise in tension with the paper's own evidence: the RETFound loss improves FID over the baseline while worsening downstream-task performance and morphological matching. The paper's conclusions should be reframed as a mixed, metric-dependent result, with the specific negative finding being that RETFound features underperform VGG perceptual features and edge-based supervision on most downstream and morphological measures.
minor comments (5)
  1. [Abstract] Several LaTeX or line-break artifacts remain in the text ('Fr\'echet', 'in-vestigated', 'autoen-coder'); these should be cleaned up before final submission.
  2. [Table 2] Table 2 uses the label 'VQGAN (RETFound)' while the text and Table 1 use 'VQ-GAN (RETFound)'; the notation should be unified.
  3. [§3.1, §4.1] Section 3.1 does not state how the 90,344 UK Biobank images were split into training, validation, and test sets, and Section 4.1 does not say whether the 'Real' row in Table 1 is computed on the same 20,000-image subset used for the synthetic comparisons; both details should be reported.
  4. [Fig. 3] The Figure 3 caption claims on the basis of a single sample that VQ-GAN (RETFound) 'struggles to replicate the optic cup'; this qualitative observation should either be tied to the quantitative optic-disc/cup features in Figure 4 or softened.
  5. [§5, ref. [33]] The citation of PeCo [33] to argue that transformer-based perceptual losses outperform pixel-based losses is not directly applicable, since PeCo is a pre-training objective rather than a frozen-feature reconstruction loss; the relevance to the present setup should be explained or the sentence revised.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the comparison is empirical and evaluation metrics are independent of training losses; confounds are acknowledged, not hidden.

full rationale

This paper makes an empirical comparison among VQ-GAN training losses: RETFound loss, perceptual loss, edge loss, and pixel-only baseline. There is no derivation chain in which a predicted quantity is defined in terms of the fitted quantity. RETFound, VGG-16, and AutoMorph are external pretrained models used as fixed feature extractors or evaluation tools; their weights are not trained or fitted on the UK Biobank data in this study. The evaluation metrics (FID, MMD, MS-SSIM, AutoMorph-based downstream prediction errors, and permutation tests on extracted retinal features) are computed from the generated samples and are not used as training objectives, so there is no fitted-input-called-prediction pattern. The central negative claim that domain-specific deep features do not improve synthesis is an inductive generalization from a comparative experiment, not a tautology. The skeptic concern that the RETFound loss and perceptual loss differ along several axes at once (architecture, number of layers, distance norm, loss weight reporting) is a confounding/overgeneralization issue about external validity, not circularity. The paper itself flags these confounds in Section 5, noting that the underperformance 'may also stem from differences in model architecture' and that 'another possible reason for the relative drop in RETFound loss performance compared to perceptual loss is the difference in the number of activation layers used.' That acknowledgment further shows the result is not presented as forced by construction. The only self-citations (references 9, 21, 22) are contextual prior work by overlapping authors and are not load-bearing for the central claim. No equation in the paper reduces to its own input, and no uniqueness theorem or ansatz is imported via self-citation. Therefore the circularity score is 0.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central empirical comparison rests on unreported loss weights, a single-layer RETFound loss, and the assumption that AutoMorph features and downstream prediction errors measure clinical fidelity.

free parameters (3)
  • Auxiliary loss weighting coefficients = not reported
    Every auxiliary loss must be weighted against pixel and GAN losses; the strength of the RETFound, edge, and perceptual terms is not stated, so the comparison is uncontrolled.
  • Layer set for RETFound loss = 1 (final transformer block)
    Only the encoder's last layer is used, while VGG perceptual loss uses five layers; this design choice is a potential confound.
  • Resizing resolution for RETFound = 224x224
    Images are resized from 256x256 to 224x224 for RETFound, possibly removing vascular detail; this is not controlled across losses.
assumptions (4)
  • domain assumption RETFound final-layer embeddings capture clinically relevant retinal morphology
    Section 3.3: the loss assumes L1 distance in RETFound latent space enforces domain-specific fidelity.
  • domain assumption AutoMorph features are a valid proxy for morphological and clinical accuracy
    Sections 3.2 and 4.3: 45 extracted features are used as ground truth for evaluating synthetic images.
  • domain assumption Train-Synthetic, Test-Real with AutoMorph features measures clinical usefulness
    Section 4.2: predictive errors on real retinal features are used to rank synthetic data quality.
  • domain assumption UK Biobank 90,344 images are representative for the comparison
    Section 3.1: all models are trained on the same subset, but no demographic analysis is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Exploring the Effectiveness of Deep Features from Domain-Specific Foundation Models in Retinal Image Synthesis." pith.science (2026). https://pith.science/paper/SF4SKGEK

@misc{pith2026250611753,
  author       = {Pith},
  title        = {Pith review of: Exploring the Effectiveness of Deep Features from Domain-Specific Foundation Models in Retinal Image Synthesis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SF4SKGEK}},
  note         = {Machine review of arXiv:2506.11753}
}
read the original abstract

The adoption of neural network models in medical imaging has been constrained by strict privacy regulations, limited data availability, high acquisition costs, and demographic biases. Deep generative models offer a promising solution by generating synthetic data that bypasses privacy concerns and addresses fairness by producing samples for under-represented groups. However, unlike natural images, medical imaging requires validation not only for fidelity (e.g., Fr\'echet Inception Score) but also for morphological and clinical accuracy. This is particularly true for colour fundus retinal imaging, which requires precise replication of the retinal vascular network, including vessel topology, continuity, and thickness. In this study, we in-vestigated whether a distance-based loss function based on deep activation layers of a large foundational model trained on large corpus of domain data, colour fundus imaging, offers advantages over a perceptual loss and edge-detection based loss functions. Our extensive validation pipeline, based on both domain-free and domain specific tasks, suggests that domain-specific deep features do not improve autoen-coder image generation. Conversely, our findings highlight the effectiveness of con-ventional edge detection filters in improving the sharpness of vascular structures in synthetic samples.

Figures

Figures reproduced from arXiv: 2506.11753 by the authors.

Figure 1
Figure 1. The model is a standard VQ-GAN with an autoencoder-decoder architecture, incorporat￾ing quantization reparameterization in the latent space and a CNN-based patch-based discrimi￾nator supervision. The proposed training strategy introduces a loss function defined in the latent space of RETFound [3] trained on 1.6 million retinal images. This encoder-decoder model, referred to as the generator, is trained using GAN los… view at source ↗
Figure 2
Figure 2. Meijering filter-based edge loss: Computes the L2 distance between vesselness scores to enforce topological correctness and vessel continuity. Note that the visualization shows a thresholded signal, while the loss operates on the continuous vesselness response. 3.5 Perceptual loss Perceptual loss is computed using a VGG-16 [30] backbone trained on ImageNet for classification. The loss is calculated as the sum of L2 … view at source ↗
Figure 3
Figure 3. Qualitative comparison of tested VQ-GAN models on a single sample. All models achieve good reconstruction, but VQ-GAN (Edge) and VQ-GAN (Perceptual) exhibit slightly sharper contrast, enhancing vessel detection, whilst VQ-GAN (RETFound) struggles to replicate the optic cup. As shown in [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Comparison of retinal features related to the optic disc and cup. Each boxplot shows the mean and standard deviation for each model. Non-grey distributions indicate samples’ overlap with the real data distribution, as supported by a permutation test (p-value > 0.05). A…
Figure 5
Figure 5. Figure 5: Comparison of retinal features related to vascular average width. Each boxplot shows the mean and standard deviation for each model. Non-grey distributions indicate samples’ over￾lap with the real data distribution, as supported by a permutation test (p-value > 0.05). …
Figure 6
Figure 6. Figure 6: Comparison of retinal features related to fractal dimension, vessel density and vascular calibre metrics. Each boxplot shows the mean and standard deviation for each model. Non-grey distributions indicate samples’ overlap with the real data distribution, as supported b…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

34 extracted references · 15 canonical work pages

  1. [1]

    Poplin et al., ‘Prediction of cardiovascular risk factors from retinal fundus photographs via deep learning’, Nat Biomed Eng, vol

    R. Poplin et al., ‘Prediction of cardiovascular risk factors from retinal fundus photographs via deep learning’, Nat Biomed Eng, vol. 2, no. 3, pp. 158–164, Feb. 2018, doi: 10.1038/s41551-018-0195-0

    work page doi:10.1038/s41551-018-0195-0 2018

  2. [2]

    C. Y. Cheung et al., ‘A deep-learning system for the assessment of cardiovas- cular disease risk via the measurement of retinal-vessel calibre’, Nat Biomed Eng, vol. 5, no. 6, 2021, doi: 10.1038/s41551-020-00626-4

    work page doi:10.1038/s41551-020-00626-4 2021

  3. [3]

    Zhou et al., ‘A foundation model for generalizable disease detection from retinal images’, Nature, vol

    Y. Zhou et al., ‘A foundation model for generalizable disease detection from retinal images’, Nature, vol. 622, no. 7981, pp. 156–163, Sep. 2023, doi: 10.1038/s41586-023-06555-x

    doi:10.1038/s41586-023-06555-x 2023

  4. [4]

    C. Y. Cheung et al., ‘A deep learning model for detection of Alzheimer’s dis- ease based on retinal photographs: a retrospective, multicentre case-control study’, Lancet Digit Health, vol. 4, no. 11, 2022, doi: 10.1016/S2589- 7500(22)00169-8

    doi:10.1016/s2589- 2022

  5. [5]

    An et al., ‘Glaucoma Diagnosis with Machine Learning Based on Optical Coherence Tomography and Color Fundus Images’, J Healthc Eng, vol

    G. An et al., ‘Glaucoma Diagnosis with Machine Learning Based on Optical Coherence Tomography and Color Fundus Images’, J Healthc Eng, vol. 2019, 2019, doi: 10.1155/2019/4061313

    work page doi:10.1155/2019/4061313 2019

  6. [6]

    Zhou et al., ‘AutoMorph: Automated Retinal Vascular Morphology Quan- tification Via a Deep Learning Pipeline’, Translational Vision Science & Technology, vol

    Y. Zhou et al., ‘AutoMorph: Automated Retinal Vascular Morphology Quan- tification Via a Deep Learning Pipeline’, Translational Vision Science & Technology, vol. 11, no. 7, p. 12, Jul. 2022, doi: 10.1167/tvst.11.7.12

    doi:10.1167/tvst.11.7.12 2022

  7. [7]

    Staal, M

    J. Staal, M. D. Abràmoff, M. Niemeijer, M. A. Viergever, and B. Van Ginneken, ‘Ridge-based vessel segmentation in color images of the retina’, IEEE Trans Med Imaging, vol. 23, no. 4, 2004, doi: 10.1109/TMI.2004.825627

    arXiv 2004

  8. [8]

    Hoover, ‘Locating blood vessels in retinal images by piecewise threshold probing of a matched filter response’, IEEE Trans Med Imaging, vol

    A. Hoover, ‘Locating blood vessels in retinal images by piecewise threshold probing of a matched filter response’, IEEE Trans Med Imaging, vol. 19, no. 3, 2000, doi: 10.1109/42.845178

    doi:10.1109/42.845178

Show all 34 references
  1. [9]

    Alloula, R

    A. Alloula, R. Mustafa, D. R. McGowan, and B. W. Papież, ‘On Biases in a UK Biobank-based Retinal Image Classification Model’, 2024. [Online]. Available: https://arxiv.org/abs/2408.02676

    2024 arXiv

  2. [10]

    Esser, R

    P. Esser, R. Rombach, and B. Ommer, ‘Taming Transformers for High-Reso- lution Image Synthesis’, 2021. [Online]. Available: https://arxiv.org/abs/2012.09841

    2021 arXiv

  3. [11]

    Rombach, A

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, ‘High-Reso- lution Image Synthesis with Latent Diffusion Models’, in Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recog- nition, 2022. doi: 10.1109/CVPR52688.2022.01042. 13

    2022

  4. [12]

    J. Ho, A. Jain, and P. Abbeel, ‘Denoising diffusion probabilistic models’, in Advances in Neural Information Processing Systems, 2020

    2020

  5. [13]

    J. Song, C. Meng, and S. Ermon, ‘DENOISING DIFFUSION IMPLICIT MODELS’, in ICLR 2021 - 9th International Conference on Learning Repre- sentations, 2021

    2021

  6. [14]

    Tudosiu et al., ‘Morphology-preserving Autoregressive 3D Generative Modelling of the Brain’, 2022, arXiv

    P.-D. Tudosiu et al., ‘Morphology-preserving Autoregressive 3D Generative Modelling of the Brain’, 2022, arXiv. doi: 10.48550/ARXIV.2209.03177

    2022 doi arXiv

  7. [15]

    X. Zhu, W. Zhang, Y. Li, L. J. O’Donnell, and F. Zhang, ‘When Diffusion MRI Meets Diffusion Model: A Novel Deep Generative Model for Diffusion MRI Generation’, 2024, arXiv. doi: 10.48550/ARXIV.2408.12897

    2024 doi arXiv

  8. [16]

    Litrico, F

    M. Litrico, F. Guarnera, M. V. Giuffrida, D. Ravì, and S. Battiato, ‘TADM: Temporally-Aware Diffusion Model for Neurodegenerative Progression on Brain MRI’, in proceedings of Medical Image Computing and Computer As- sisted Intervention – MICCAI 2024, Springer Nature Switzerlan...

    2024

  9. [17]

    N. Konz, Y. Chen, H. Dong, and M. A. Mazurowski, ‘Anatomically-Control- lable Medical Image Generation with Segmentation-Guided Diffusion Mod- els’, in proceedings of Medical Image Computing and Computer Assisted In- tervention – MICCAI 2024, Springer Nature Switzerland, Oct. 2024

    2024

  10. [18]

    Doerrich, F

    S. Doerrich, F. Di Salvo, and C. Ledig, ‘Self-supervised Vision Transformer are Scalable Generative Models for Domain Generalization’, in proceedings of Medical Image Computing and Computer Assisted Intervention – MICCAI 2024, Springer Nature Switzerland, Oct. 2024

    2024

  11. [19]

    Sturm, L

    M. Sturm, L. Cerrone, and F. A. Hamprecht, ‘SynCellFactory: Generative Data Augmentation for Cell Tracking’, in proceedings of Medical Image Computing and Computer Assisted Intervention – MICCAI 2024, Springer Nature Switzerland, Oct. 2024

    2024

  12. [20]

    Zhang et al., ‘LeFusion: Controllable Pathology Synthesis via Lesion-Fo- cused Diffusion Models’, 2024, arXiv

    H. Zhang et al., ‘LeFusion: Controllable Pathology Synthesis via Lesion-Fo- cused Diffusion Models’, 2024, arXiv. doi: 10.48550/ARXIV.2403.14066

    2024 doi arXiv

  13. [21]

    Bradbury, K

    R. Bradbury, K. A. Vallis, and B. W. Papiez, ‘Paired Diffusion: Generation of related, synthetic PET-CT-Segmentation scans using Linked Denoising Diffu- sion Probabilistic Models’, Proceedings - International Symposium on Bio- medical Imaging, Mar. 2024, doi: 10.1109/ISBI56570...

    2024

  14. [22]

    Zhang, F

    L. Zhang, F. Wu, K. Bronik, and B. W. Papiez, ‘DiffuSeg: Domain-driven Diffusion for Medical Image Segmentation’, IEEE J Biomed Health Inform, 2025, doi: 10.1109/JBHI.2025.3526806

    2025

  15. [23]

    Hou et al., ‘FundusGAN: A Hierarchical Feature-Aware Generative Framework for High-Fidelity Fundus Image Generation’, Mar

    Q. Hou et al., ‘FundusGAN: A Hierarchical Feature-Aware Generative Framework for High-Fidelity Fundus Image Generation’, Mar. 2025, Ac- cessed: Apr. 05, 2025. [Online]. Available: https://arxiv.org/abs/2503.17831v1

    2025 arXiv

  16. [24]

    S. Ahn, S. J. Song, and J. Shin, ‘FundusGAN: Fundus image synthesis based on semi-supervised learning’, Biomed Signal Process Control, vol. 86, p. 105289, Sep. 2023, doi: 10.1016/J.BSPC.2023.105289

    2023

  17. [25]

    Zhang, P

    R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang, ‘The Unreason- able Effectiveness of Deep Features as a Perceptual Metric’, in 2018 14 IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018, pp. 586–595. doi: 10.1109/CVPR.2018.00068

    2018

  18. [26]

    K. He, X. Zhang, S. Ren, and J. Sun, ‘Deep Residual Learning for Image Recognition’, Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, vol. 2016-December, pp. 770–778, Dec. 2015, doi: 10.1109/CVPR.2016.90

    2016 doi

  19. [27]

    van den Oord, O

    A. van den Oord, O. Vinyals, and K. Kavukcuoglu, ‘Neural Discrete Repre- sentation Learning’, 2018. [Online]. Available: https://arxiv.org/abs/1711.00937

    2018 arXiv

  20. [28]

    Dosovitskiy et al., ‘An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale’, 2021

    A. Dosovitskiy et al., ‘An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale’, 2021. [Online]. Available: https://arxiv.org/abs/2010.11929

    2021 arXiv

  21. [29]

    Meijering, M

    E. Meijering, M. Jacob, J. C. F. Sarria, P. Steiner, H. Hirling, and M. Unser, ‘Design and Validation of a Tool for Neurite Tracing and Analysis in Fluores- cence Microscopy Images’, Cytometry Part A, vol. 58, no. 2, 2004, doi: 10.1002/cyto.a.20022

    2004 doi

  22. [30]

    Simonyan and A

    K. Simonyan and A. Zisserman, ‘Very Deep Convolutional Networks for Large-Scale Image Recognition’, 3rd International Conference on Learning Representations, ICLR 2015 - Conference Track Proceedings, Sep. 2014, Ac- cessed: Apr. 06, 2025. [Online]. Available: https://arxiv.org/...

    2015 arXiv

  23. [31]

    Szegedy, V

    C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna, ‘Rethinking the Inception Architecture for Computer Vision’, in Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition ,

  24. [32]

    Jordon et al., ‘Synthetic Data -- what, why and how?’, May 2022

    J. Jordon et al., ‘Synthetic Data -- what, why and how?’, May 2022

    2022

  25. [33]

    Dong et al., ‘PeCo: Perceptual Codebook for BERT Pre-training of Vision Transformers’, Proceedings of the AAAI Conference on Artificial Intelligence, vol

    X. Dong et al., ‘PeCo: Perceptual Codebook for BERT Pre-training of Vision Transformers’, Proceedings of the AAAI Conference on Artificial Intelligence, vol. 37, no. 1, pp. 552–560, Jun. 2023, doi: 10.1609/AAAI.V37I1.25130

    2023 doi

  26. [2016]

    doi: 10.1109/CVPR.2016.308

    2016 doi

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.