Pith. sign in

REVIEW 3 major objections 5 minor 40 references

Efficacy of Image Similarity as a Metric for Augmenting Small Dataset Retinal Image Segmentation

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Lower FID predicts larger segmentation gains, but only when synthetic images are sufficiently dissimilar — and the trend is log-normal, not linear.

desk verdict A careful small study whose abstract overstates the evidence for separate log-normal FID-DSC trends; the data support a weaker qualitative claim. read the letter →

arxiv 2507.04862 v3 pith:AWDCFERM submitted 2025-07-07 eess.IV cs.CV

Thomas Wallace , Ik Siong Heng , Senad Subasic , Chris Messenger This is my paper
classification eess.IVcs.CV
keywords syntheticdataaugmentationFréchetInceptionDistanceretinalOCTsegmentationU-NetPGGANsmalldatasetintraretinalfluidBayesfactor
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 tests whether the Fréchet Inception Distance (FID) — a standard measure of how similar two image datasets are — can predict how much synthetic images improve a U-Net segmentation model trained on a small retinal OCT dataset. Using PGGAN-generated synthetic images and standard augmentations (flips, blur, noise) to augment Diabetes-related Macular Edema intraretinal fluid segmentation, the authors find that high-FID (dissimilar) datasets barely help, while lower-FID datasets produce significant and robust gains. They report that the improvement grows as FID decreases but follows separate log-normal curves for synthetic versus standard augmentations, with synthetic data outperforming standard augmentations at the same FID. The paper concludes that FID is a useful but imperfect indicator: images that are too similar may stop helping, so 'sufficient dissimilarity' may be required for augmentation to work.

What carries the argument

The load-bearing object is the FID itself: the Wasserstein-2 distance between two 2048-dimensional multivariate Gaussians estimated from Inception network feature vectors of the real and augmented image sets. It serves as the independent variable measuring dataset similarity. The response is the mean Dice Similarity Coefficient of the U-Net on withheld test data, and model selection uses the Bayes factor among constant, linear, log-normal, and inverse curves. The log-normal form $h(x) = \mathrm{DSC}_0 + A \exp\left(-\frac{1}{2}\left(\frac{\ln x - \mu}{\sigma}\right)^2\right)$ carries the paper's main conclusion that improvement rises then falls as FID decreases.

What would settle it

Re-run the same augmentation experiments but compute FID with a feature extractor trained on retinal OCT images (or use a perceptual similarity metric tuned to medical imaging) and check whether the two log-normal trends and the drop at low FID persist; if the relationship changes form or disappears, the ImageNet-based FID was carrying the result. Also, test synthetic datasets with FID ≤ 50 to see whether the predicted downward leg of the log-normal actually occurs.

Watch

Extended reading notes

Core claim

The paper's central claim is that FID, computed between the training dataset and an augmentation dataset, is predictive of the average Dice improvement a U-Net obtains from that augmentation, but not in a simple monotone way. The authors show that dissimilar (high FID) augmentation datasets do not improve segmentation significantly, and that as FID decreases the augmentation contributes significant, robust improvements. Bayesian model comparison across four candidate curves (constant, linear, log-normal, inverse) favours a log-normal relationship between FID and DSC improvement in nearly every condition tested, and the data indicate that synthetic and standard augmentations follow two different log-normal trends. The paper interprets the low-FID downward leg of the log-normal as possible data memorization by the PGGAN: very similar synthetic images carry little novel information. It concludes that FID should not be used alone as a universal metric for augmentation effectiveness.

Load-bearing premise

The entire analysis treats FID values computed from an Inception network trained on ordinary photographs as the correct measure of similarity for medical OCT images; if those features miss what makes two OCT datasets clinically similar, the observed FID-DSC trends could be artifacts of that specific embedding.

Editorial extensions

If this is right

  • FID can be used as a screening tool: synthetic augmentations with high FID relative to the training set are unlikely to help and may be skipped.
  • The separate log-normal trends imply that comparing synthetic and standard augmentation strategies by FID alone is misleading; the generation method matters.
  • A log-normal FID-DSC relationship means there is an optimal similarity range, and datasets that are too similar (very low FID) can yield less improvement.
  • When augmenting small medical datasets, researchers should report FID alongside DSC gains, since both determine whether the augmentation was worthwhile.
  • The finding that synthetic data outperforms standard augmentations at equal FID suggests generative-model-based augmentation is a particularly efficient route for small retinal datasets.

Reading between the lines

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

  • If the ImageNet-trained Inception features are the reason FID behaves this way, then replacing them with a domain-specific (OCT-trained) embedding should either sharpen or dissolve the two log-normal trends; this is directly testable with the same augmentation sets.
  • The 'sufficient dissimilarity' requirement suggests a novelty-information tradeoff: extremely similar images add redundant information, so an augmentation's value may peak at a FID that balances familiarity with novelty.
  • The same experimental protocol could be applied to diffusion-model-generated OCT images; the paper's separation of standard and synthetic trends predicts that each generative family will have its own FID-effectiveness curve, which would further weaken the case for FID as a universal metric.
  • A practical extension: plot DSC improvement against FID for a new pathology or imaging device first, then pick augmentation datasets from the rising part of the curve; the paper's data imply that blindly adding the lowest-FID synthetic images is not always optimal.
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

3 major / 5 minor

Summary. The paper investigates whether the Fréchet Inception Distance (FID) between a training set and an augmentation set predicts improvements in U-Net segmentation performance for Diabetes-related Macular Edema intraretinal fluid in retinal OCT B-scans. It reports three experiments: (i) the effect of increasing the number of standard or PGGAN-generated synthetic augmentation images on Dice Similarity Coefficient (DSC); (ii) how augmentation datasets with different FIDs change the number-vs-DSC relationship; and (iii) a model-selection analysis using Bayes factors to compare constant, linear, log-normal-shaped, and inverse models of the FID-DSC relationship. The paper's main positive result is that lower-FID augmentation data tend to produce larger and more robust DSC improvements, while high-FID data provide little benefit. Its most distinctive claim is that synthetic and standard augmentations follow separate log-normal trends, with synthetic data more effective. The authors candidly list limitations: a single segmentation task, PGGAN-only generation, ImageNet-trained Inception features, limited numbers of samples, and no external validation.

Significance. If the FID-DSC relationship is real, the paper provides useful practical guidance for selecting augmentation data in low-data medical imaging and contributes to the ongoing discussion of whether FID is a reliable quality metric for synthetic medical images. Strengths include the independent measurement of the independent and dependent variables (so the headline correlation is not an artifact of circular definition), a principled Bayes-factor comparison framework, explicit reporting of standard errors and outlier behavior, and an unusually candid discussion of computational and domain limitations. The most novel claim—separate log-normal trends for synthetic and standard augmentations—is nevertheless overstated relative to the evidence in Table 3 and is not formally tested against a combined model, which limits the reliability of the conclusions until revised.

major comments (3)
  1. [Abstract; Section 6.2, Table 3] The abstract states there is "significant evidence" that synthetic and standard augmentations follow separate log-normal trends, but Table 3 does not support this for standard augmentations. For Standard +25%, the log-normal model h and the constant model f have identical log-evidence (7.661 vs. 7.661); at +50% the preference is negligible (6.906 vs. 6.87); and at +75% the linear model g is preferred over h (4.928 vs. 4.204). Only at +99% is h clearly preferred. Section 7 itself concedes that the preference is only substantial in one standard case. The abstract and conclusion should be qualified to state that the log-normal evidence is strong for synthetic data, while for standard data it is generally not preferred over simpler models.
  2. [Section 6.1, Eqs. (5)–(10)] The "separate trends" claim is never formally tested. All four models in Eqs. (5a)–(5d) are fit independently to the synthetic and standard datasets, and the Bayes factors in Table 3 compare models within each augmentation type. No model is defined that pools the two types or shares parameters, so no evidence is computed for "one common trend for both types" versus "two separate trends." The different best-fit parameters in Table B.1 are necessary but not sufficient to establish separation; a formal comparison (e.g., a common-parameter log-normal model versus type-specific parameters) is required before claiming significant evidence of separate trends.
  3. [Abstract; Appendix B] The abstract's claim that synthetic data "prove more effective" than standard augmentations is not established by the fitted parameters. In Table B.1 at +99%, A_standard = 0.079 while A_synthetic = 0.057, and the differing values of mu and sigma mean the ordering of the two curves depends on FID. The paper reports no uncertainty intervals for the fitted parameters and no statistical comparison of synthetic versus standard performance at matched FID values, so the relative-effectiveness statement should be softened or formally tested.
minor comments (5)
  1. [Section 5.2] The sentence "we see that low FID images do not improve the model significantly" appears to be a typo for "high FID", since the next sentence and Figure 5 show that the lowest-FID (rightmost) dataset produces the largest improvement.
  2. [Eq. (9)] The notation "1QNθ i Ri" in Equation (9) should likely read 1/(prod_i R_i); please correct this typographical issue.
  3. [Table 3 and Section 6.2] Please specify whether the log-evidences in Table 3 are natural logs or base-10 logs. The text says log10 K is obtained by directly subtracting log-evidences, which is only valid if the table entries are already base-10 values.
  4. [Figure 6 caption] The caption states a log-normal model is chosen in all cases "in accordance with the evidences in Table 3," but for Standard +75% the linear model has the highest log-evidence, so the statement overstates the support for the log-normal fits.
  5. [Section 6.2] Consider replacing "log-normal model" with a term like "log-Gaussian curve", since h(x) in Eq. (5c) is a Gaussian bump in log-FID rather than a log-normal probability density.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: FID and DSC are measured independently, and the log-normal model fitting is descriptive rather than definitional.

full rationale

The paper's derivation chain is self-contained and does not reduce to its inputs. FID is defined in Eq. (3) from pretrained Inception feature statistics, and DSC is defined in Eq. (4) from mask overlap; the two quantities are measured independently. The augmentation datasets are generated by a PGGAN and the U-Nets are trained and evaluated separately, so no equation defines DSC in terms of FID by construction. The log-normal model h(x) in Eq. (5c) is fitted to the observed (FID, DSC) pairs, and the Bayes-factor model selection in Table 3 is a standard posterior comparison on the same data, not a prediction of a quantity that was used as an input to the fit. The paper's stronger conclusion about 'separate log-normal trends' is statistically fragile, as Table 3 shows the log-normal model is only marginally preferred for standard augmentations and no combined model is tested, but that is an evidence-strength and overclaim concern, not circularity. The ImageNet/Inception limitation is explicitly acknowledged in Section 7 as a potential validity threat, not hidden or repackaged. No load-bearing self-citation appears; cited prior studies are external. The finding is therefore a normal empirical result with no definitional or fitted-input circularity.

Assumptions & free parameters 8 free parameters · 4 assumptions · 0 invented entities

The paper introduces no invented entities. Its main assumptions are statistical (a normal likelihood and chosen uniform priors for Bayes factor computation), domain-specific (ImageNet-based FID applied to OCT scans), and external-validity related (test set from the same source dataset). The paper itself flags most of these in Section 7.

free parameters (8)
  • Synthetic +25% log-normal (A, mu, sigma) = 0.040, 4.617, 0.634
    Best-fit parameters from Table B.1 used to draw the +25% synthetic curve in Fig. 6; supports the log-normal trend claim.
  • Standard +25% log-normal (A, mu, sigma) = 0.014, 3.146, 2.600
    Best-fit parameters from Table B.1 used to draw the +25% standard curve in Fig. 6; supports the log-normal trend claim.
  • Synthetic +50% log-normal (A, mu, sigma) = 0.048, 4.687, 0.473
    Best-fit parameters from Table B.1 used to draw the +50% synthetic curve in Fig. 6; supports the log-normal trend claim.
  • Standard +50% log-normal (A, mu, sigma) = 0.023, 2.015, 2.311
    Best-fit parameters from Table B.1 used to draw the +50% standard curve in Fig. 6; supports the log-normal trend claim.
  • Synthetic +75% log-normal (A, mu, sigma) = 0.055, 4.535, 0.495
    Best-fit parameters from Table B.1 used to draw the +75% synthetic curve in Fig. 6; supports the log-normal trend claim.
  • Standard +75% log-normal (A, mu, sigma) = 0.035, 0.785, 1.785
    Best-fit parameters from Table B.1 used to draw the +75% standard curve in Fig. 6; supports the log-normal trend claim.
  • Synthetic +99% log-normal (A, mu, sigma) = 0.057, 4.552, 0.467
    Best-fit parameters from Table B.1 used to draw the +99% synthetic curve in Fig. 6; supports the log-normal trend claim.
  • Standard +99% log-normal (A, mu, sigma) = 0.079, 3.021, 0.535
    Best-fit parameters from Table B.1 used to draw the +99% standard curve in Fig. 6; supports the log-normal trend claim.
assumptions (4)
  • domain assumption Average DSC values follow a normal distribution around the model prediction.
    Assumed in the likelihood in Eq. 10; the paper acknowledges longer tails toward low DSC and outliers in Section 6.2 and Fig. 7.
  • ad hoc to paper Uniform priors on model parameters (Eq. 6) are reasonable and fair.
    The Bayes factors in Table 3 are conditional on these priors; Section 7 states the preference is derived from the chosen priors.
  • domain assumption FID from an ImageNet-trained Inception network is a meaningful similarity metric for retinal OCT images.
    The entire study uses FID as the independent variable; Section 7 flags the ImageNet training as a possible weakness.
  • domain assumption Results on the withheld test set will transfer to other acquisition protocols or institutions.
    The paper lists lack of external validation as a limitation in Section 7; the practical claim about FID usefulness assumes some generalizability.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Efficacy of Image Similarity as a Metric for Augmenting Small Dataset Retinal Image Segmentation." pith.science (2026). https://pith.science/paper/AWDCFERM

@misc{pith2026250704862,
  author       = {Pith},
  title        = {Pith review of: Efficacy of Image Similarity as a Metric for Augmenting Small Dataset Retinal Image Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AWDCFERM}},
  note         = {Machine review of arXiv:2507.04862}
}
read the original abstract

Synthetic images are an option for augmenting limited medical imaging datasets to improve the performance of various machine learning models. A common metric for evaluating synthetic image quality is the Fr\'echet Inception Distance (FID) which measures the similarity of two image datasets. In this study we evaluate the relationship between this metric and the improvement which synthetic images, generated by a Progressively Growing Generative Adversarial Network (PGGAN), grant when augmenting Diabetes-related Macular Edema (DME) intraretinal fluid segmentation performed by a U-Net model with limited amounts of training data. We find that the behaviour of augmenting with standard and synthetic images agrees with previously conducted experiments. Additionally, we show that dissimilar (high FID) datasets do not improve segmentation significantly. As FID between the training and augmenting datasets decreases, the augmentation datasets are shown to contribute to significant and robust improvements in image segmentation. Finally, we find that there is significant evidence to suggest that synthetic and standard augmentations follow separate log-normal trends between FID and improvements in model performance, with synthetic data proving more effective than standard augmentation techniques. Our findings show that more similar datasets (lower FID) will be more effective at improving U-Net performance, however, the results also suggest that this improvement may only occur when images are sufficiently dissimilar.

Figures

Figures reproduced from arXiv: 2507.04862 by the authors.

Figure 1
Figure 1. Example of a PGGAN network showing the gradual addition of new [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Example of how training data was generated by combining the [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. An example of a U-Net architecture showing the encoding con [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: The Results of Section 4 where we increase the number of images [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Results of Section 5 showing the mean and standard error of the [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: Results of Section 6 showing the relationship between FID and DSC [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 7
Figure 7. Figure 7: An example distribution for synthetic data with 100 samples show [PITH_FULL_IMAGE:figures/full_fig_p017_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

40 extracted references · 19 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in ":" * " " * FUNCTION f...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in ":" * " " * FUNCTION f...

  3. [3]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in ":" * " " * FUNCTION f...

  4. [4]

    , author Al-Mousa, A.A

    author Alqudah, R. , author Al-Mousa, A.A. , author Hashyeh, Y.A. , author Alzaibaq, O.Z. , year 2023 . title A systemic comparison between using augmented data and synthetic data as means of enhancing wafermap defect classification . journal Computers in Industry volume 145 , pages 103809 . https://www.sciencedirect.com/science/article/pii/S0166361522002...

    arXiv 2023

  5. [5]

    Synthetic Image Data for Deep Learning

    author Anderson, J.W. , author Ziolkowski, M. , author Kennedy, K. , author Apon, A.W. , year 2022 . title Synthetic image data for deep learning . https://arxiv.org/abs/2212.06232, http://arxiv.org/abs/2212.06232 arXiv:2212.06232

    work page Pith review arXiv 2022

  6. [6]

    , author Chintala, S

    author Arjovsky, M. , author Chintala, S. , author Bottou, L. , year 2017 . title Wasserstein GAN . https://arxiv.org/abs/1701.07875, :10.48550/ARXIV.1701.07875

    arXiv doi:10.48550/arxiv.1701.07875 2017

  7. [7]

    , author Javaid, U

    author Barragán-Montero, A. , author Javaid, U. , author Valdés, G. , author Nguyen, D. , author Desbordes, P. , author Macq, B. , author Willems, S. , author Vandewinckele, L. , author Holmström, M. , author Löfman, F. , author Michiels, S. , author Souris, K. , author Sterpin, E. , author Lee, J.A. , year 2021 . title Artificial intelligence and machine...

    2021

  8. [8]

    , author Arora, S

    author Bhatia, K. , author Arora, S. , author Tomar, R. , year 2016 . title Diagnosis of diabetic retinopathy using machine learning classification algorithm , in: booktitle 2016 2nd International Conference on Next Generation Computing Technologies (NGCT) , pp. pages 347--351 . :10.1109/NGCT.2016.7877439

    arXiv 2016

Show all 40 references
  1. [9]

    , author van der Schaar, M

    author van Breugel, B. , author van der Schaar, M. , year 2023 . title Beyond privacy: Navigating the opportunities and challenges of synthetic data . https://arxiv.org/abs/2304.03722, http://arxiv.org/abs/2304.03722 arXiv:2304.03722

    2023 arXiv

  2. [10]

    , author Joshi, N

    author Burlina, P. , author Joshi, N. , author Paul, W. , author Pacheco, K.D. , author Bressler, N.M. , year 2021 . title Addressing Artificial Intelligence Bias in Retinal Diagnostics . journal Translational Vision Science & Technology volume 10 , pages 13--13 . https://doi....

    2021

  3. [11]

    , author Min, H

    author Chlap, P. , author Min, H. , author Vandenberg, N. , author Dowling, J. , author Holloway, L. , author Haworth, A. , year 2021 . title A review of medical image data augmentation techniques for deep learning applications . journal Journal of Medical Imaging and Radiatio...

    2021

  4. [12]

    , author Schmidt, V

    author Courty, B. , author Schmidt, V. , author Luccioni, S. , author Kamal, G. , author Coutarel, M. , author Feld, B. , author Lecourt, J. , author Connell, L. , author Saboni, A. , author Inimaz , author supatomic , author Léval, M. , author Blanche, L. , author Cruveiller,...

    2024

  5. [13]

    , author Tran, R

    author Das, H.P. , author Tran, R. , author Singh, J. , author Yue, X. , author Tison, G. , author Sangiovanni-Vincentelli, A. , author Spanos, C.J. , year 2022 . title Conditional synthetic data generation for robust machine learning applications with limited pandemic data , ...

    2022

  6. [14]

    title Greenhouse gas reporting: conversion factors 2023

    author Department for Energy Security and Net Zero , year 2023 . title Greenhouse gas reporting: conversion factors 2023

    2023

  7. [15]

    , author Landau, B

    author Dowson, D. , author Landau, B. , year 1982 . title The F réchet distance between multivariate normal distributions . journal Journal of Multivariate Analysis volume 12 , pages 450--455 . https://www.sciencedirect.com/science/article/pii/0047259X8290077X, :https://doi.or...

    1982

  8. [16]

    , author Klang, E

    author Frid-Adar, M. , author Klang, E. , author Amitai, M. , author Goldberger, J. , author Greenspan, H. , year 2018 . title Synthetic data augmentation using GAN for improved liver lesion classification . http://arxiv.org/abs/1801.02385 arXiv:1801.02385

    2018 arXiv

  9. [17]

    , author Anjos, A

    author Galdran, A. , author Anjos, A. , author Dolz, J. , author Chakor, H. , author Lombaert, H. , author Ayed, I.B. , year 2022 . title State-of-the-art retinal vessel segmentation with minimalistic models . journal Scientific Reports volume 12 , pages 6174 . https://doi.org...

    2022 doi

  10. [18]

    , author Shung, D.L

    author Giuffr \`e , M. , author Shung, D.L. , year 2023 . title Harnessing the power of synthetic data in healthcare: innovation, application, and privacy . journal npj Digital Medicine volume 6 , pages 186 . https://doi.org/10.1038/s41746-023-00927-3, :10.1038/s41746-023-00927-3

    2023 doi

  11. [19]

    , author Pouget-Abadie, J

    author Goodfellow, I.J. , author Pouget-Abadie, J. , author Mirza, M. , author Xu, B. , author Warde-Farley, D. , author Ozair, S. , author Courville, A. , author Bengio, Y. , year 2014 . title Generative adversarial networks . https://arxiv.org/abs/1406.2661, :10.48550/ARXIV....

    2014 doi arXiv

  12. [20]

    , author Ramsauer, H

    author Heusel, M. , author Ramsauer, H. , author Unterthiner, T. , author Nessler, B. , author Hochreiter, S. , year 2017 . title GAN s trained by a two time-scale update rule converge to a local N ash equilibrium , in: booktitle Proceedings of the 31st International Conferenc...

    2017

  13. [21]

    , author Hong, Y.J

    author Jeong, Y. , author Hong, Y.J. , author Han, J.H. , year 2022 . title Review of machine learning applications using retinal fundus images . journal Diagnostics volume 12 . https://www.mdpi.com/2075-4418/12/1/134, :10.3390/diagnostics12010134

    2022 doi

  14. [22]

    , author Aila, T

    author Karras, T. , author Aila, T. , author Laine, S. , author Lehtinen, J. , year 2017 . title Progressive growing of GAN s for improved quality, stability, and variation . https://arxiv.org/abs/1710.10196, :10.48550/ARXIV.1710.10196

    2017 doi arXiv

  15. [23]

    , author Laine, S

    author Karras, T. , author Laine, S. , author Aila, T. , year 2021 . title A Style-Based Generator Architecture for Generative Adversarial Networks . journal IEEE Transactions on Pattern Analysis & Machine Intelligence volume 43 , pages 4217--4228 . https://doi.ieeecomputersoc...

    2021

  16. [24]

    , author Raftery, A.E

    author Kass, R.E. , author Raftery, A.E. , year 1995 . title Bayes factors . journal Journal of the American Statistical Association volume 90 , pages 773--795 . http://www.jstor.org/stable/2291091

    1995

  17. [25]

    , author Zhang, K

    author Kermany, D. , author Zhang, K. , author Goldbaum, M. , year 2018 . title Labeled optical coherence tomography ( OCT ) and chest x-ray images for classification . howpublished Mendeley Data, V2, doi: 10.17632/rscbjbr9sj.2

    2018 doi

  18. [26]

    , author Kim, J.Y

    author Kong, H.J. , author Kim, J.Y. , author Moon, H.M. , author Park, H.C. , author Kim, J.W. , author Lim, R. , author Woo, J. , author Fakhri, G.E. , author Kim, D.W. , author Kim, S. , year 2022 . title Automation of generative adversarial network-based synthetic data-aug...

    2022 doi

  19. [27]

    , author Rachitskaya, A.V

    author Lim, J.I. , author Rachitskaya, A.V. , author Hallak, J.A. , author Gholami, S. , author Alam, M.N. , year 2024 . title Artificial intelligence for retinal diseases . journal Asia-Pacific Journal of Ophthalmology volume 13 , pages 100096 . https://www.sciencedirect.com/...

    2024

  20. [28]

    , author Joshi, N

    author Pekala, M. , author Joshi, N. , author Liu, T.A. , author Bressler, N. , author DeBuc, D.C. , author Burlina, P. , year 2019 . title Deep learning based retinal OCT segmentation . journal Computers in Biology and Medicine volume 114 , pages 103445 . https://www.scienced...

    2019

  21. [29]

    , author Fischer, P

    author Ronneberger, O. , author Fischer, P. , author Brox, T. , year 2015 . title U-net: Convolutional networks for biomedical image segmentation . http://arxiv.org/abs/1505.04597 arXiv:1505.04597

    2015 arXiv

  22. [30]

    , author Batouty, N.M

    author Saleh, G.A. , author Batouty, N.M. , author Haggag, S. , author Elnakib, A. , author Khalifa, F. , author Taher, F. , author Mohamed, M.A. , author Farag, R. , author Sandhu, H. , author Sewelam, A. , author El-Baz, A. , year 2022 . title The role of medical image modal...

    2022 doi

  23. [31]

    , year 2020

    author Seitzer, M. , year 2020 . title pytorch-fid: FID Score for PyTorch . howpublished https://github.com/mseitzer/pytorch-fid . note Version 0.3.0

    2020

  24. [32]

    , author Khoshgoftaar, T.M

    author Shorten, C. , author Khoshgoftaar, T.M. , year 2019 . title A survey on image data augmentation for deep learning . journal Journal of Big Data volume 6 , pages 60 . https://doi.org/10.1186/s40537-019-0197-0, :10.1186/s40537-019-0197-0

    2019 doi

  25. [33]

    , author Khanna, M

    author Singh, L.K. , author Khanna, M. , author Thawkar, S. , author Singh, R. , year 2024 . title Deep-learning based system for effective and automatic blood vessel segmentation from retinal fundus images . journal Multimedia Tools and Applications volume 83 , pages 6005--60...

    2024 doi

  26. [34]

    , author Larsson, M

    author Usman Akbar, M. , author Larsson, M. , author Blystad, I. , author Eklund, A. , year 2024 . title Brain tumor segmentation using synthetic mr images - a comparison of gans and diffusion models . journal Scientific Data volume 11 , pages 259 . https://doi.org/10.1038/s41...

    2024 doi

  27. [35]

    , author Othmani, A

    author Walsh, J. , author Othmani, A. , author Jain, M. , author Dev, S. , year 2022 . title Using u-net network for efficient brain tumor segmentation in MRI images . journal Healthcare Analytics volume 2 , pages 100098 . https://www.sciencedirect.com/science/article/pii/S277...

    2022

  28. [36]

    , author Sui, R

    author Wei, X. , author Sui, R. , year 2023 . title A review of machine learning algorithms for retinal cyst segmentation on optical coherence tomography . journal Sensors volume 23 . https://www.mdpi.com/1424-8220/23/6/3144, :10.3390/s23063144

    2023 doi

  29. [37]

    , author Yuan, S

    author Xu, D. , author Yuan, S. , author Zhang, L. , author Wu, X. , year 2018 a. title Fair GAN : Fairness-aware generative adversarial networks . http://arxiv.org/abs/1805.11202 arXiv:1805.11202

    2018 arXiv

  30. [38]

    , author Huang, G

    author Xu, Q. , author Huang, G. , author Yuan, Y. , author Guo, C. , author Sun, Y. , author Wu, F. , author Weinberger, K. , year 2018 b. title An empirical study on evaluation metrics of generative adversarial networks . https://arxiv.org/abs/1806.07755, :10.48550/ARXIV.1806.07755

    2018 doi arXiv

  31. [39]

    , author Choi, J.Y

    author Yoo, T.K. , author Choi, J.Y. , author Kim, H.K. , year 2021 . title Feasibility study to improve deep learning in OCT diagnosis of rare retinal diseases with few-shot classification . journal Med Biol Eng Comput volume 59 , pages 401--415

    2021

  32. [40]

    , author Warfield, S.K

    author Zou, K.H. , author Warfield, S.K. , author Bharatha, A. , author Tempany, C.M. , author Kaus, M.R. , author Haker, S.J. , author Wells, W.M. , author Jolesz, F.A. , author Kikinis, R. , year 2004 . title Statistical validation of image segmentation quality based on a sp...

    2004 doi

Pith tools

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