Pith. sign in

REVIEW 4 major objections 3 minor 27 references

Self-supervised denoising of visual field data improves detection of glaucoma progression

T0 review · 4 major / 3 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read The paper claims that appending a categorical p-value at every visual field location to a masked autoencoder's input improves the smoothing of visual field data, so pointwise linear regression detects more progressing eyes and predicts…

desk verdict A clinically motivated denoising comparison that fails on its own central claim: the paper's simulation data contradict the p-value benefit, but the MAE-vs-VAE comparison is worth a second look. read the letter →

arxiv 2411.12146 v1 pith:FBTQKHLB submitted 2024-11-19 eess.IV cs.CVcs.LG

Sean Wu , Jun Yu Chen , Vahid Mohammadzadeh , Sajad Besharati , Jaewon Lee , Kouros Nouri-Mahdavi , Joseph Caprioli , Zhe Fei
show 1 more author
Fabien Scalzo
This is my paper · ORCID
classification eess.IVcs.CVcs.LG
keywords glaucomaprogressionvisualfielddenoisingmaskedautoencodervariationalpointwiselinearregressionp-valueencodingself-supervisedlearningperimetry
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

Glaucoma is monitored with repeated visual field tests, but the measurements are noisy enough that progression algorithms can miss or lag real worsening. The paper argues that a self-supervised masked autoencoder — a network trained on 4,232 patients' visual field exams to reconstruct randomly zeroed-out locations of the 24-2 field — smooths the data better than raw fields or a variational autoencoder, and that appending a categorical p-value at every location makes the smoothing even more useful. In the clinical cohort, the p-value-augmented masked autoencoder flagged 21.21% of eyes as progressing by pointwise linear regression, versus 19.62% for the masked autoencoder alone, and shortened the average conversion time from 7.20 years to 7.01 years. If the claim holds, the denoised visual field could be used as a preprocessing stage in front of existing progression algorithms, improving detection without changing the downstream clinical method.

What carries the argument

The load-bearing mechanism is a masked autoencoder built from feedforward layers. During training, ten of the 52 visual field locations are zeroed out, the 52 sensitivity values are concatenated with a 52-dimensional one-hot vector of categorical p-values, and the resulting 104-dimensional input is compressed to a 16-unit bottleneck and decoded back to 52 values; the mean squared error against the pre-masking values forces the network to infer the missing locations from the surrounding field and the p-value pattern. At inference no locations are masked, so the trained network acts as a denoiser, with the p-value vector carrying per-location information about deviation from normal into the latent representation. The same pipeline is run with a variational autoencoder as the comparison arm.

What would settle it

Recompute the PLR and time-to-progression comparison using only the held-out test eyes, and check whether MAE+p-values still beats MAE; if the 1.59 percentage-point gain and 0.19-year earlier conversion disappear, the p-value benefit is an artifact of training-data leakage.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that masking out visual field locations during training and including a per-location p-value in the input produces a denoised visual field better matched to the task of detecting glaucomatous change. The masked autoencoder with p-values detected 21.21% of eyes as progressing under pointwise linear regression, compared with 19.62% for the masked autoencoder without p-values and 14.48% for raw visual field data, and it reduced the mean conversion time to 7.01 years from 7.20 years (masked autoencoder) and 7.80 years (raw data). The paper reads this as evidence that the categorical p-value acts as supplementary side information that helps the network encode how far each location deviates from normal, so the reconstructed field retains the true signal while dropping measurement noise. The proposed role of the denoiser is therefore a front-end for established progression algorithms such as PLR, MD, and GRI.

Load-bearing premise

The reported gains rest on the assumption that progression statistics were computed only on eyes the denoisers never saw during training; the Methods do not explicitly state that, so if training eyes were included the improvements could be in-sample artifacts.

Editorial extensions

If this is right

  • If the claim holds, adding per-location p-values is a practical upgrade to visual field denoising, because the p-value map is already exported with each visual field exam and no new measurement is required.
  • The denoised output plugs into pointwise linear regression without changing the regression itself, so existing glaucoma-monitoring software could adopt the preprocessing without replacing its progression metric.
  • The comparison with the variational autoencoder suggests that reconstructing missing locations is a stronger inductive bias for visual field smoothing than latent probabilistic compression, at least for this dataset and these progression metrics.
  • The p-value benefit is visible in PLR but not uniformly in MD or GRI, so the choice of denoised input may need to be tied to the downstream progression method rather than applied globally.

Reading between the lines

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

  • A natural follow-up would replace the binned one-hot p-values with a continuous per-location noise or deviation estimate; if the mechanism is uncertainty encoding, the continuous input could preserve or extend the gain without binning artifacts.
  • The masking trick could be applied across visits rather than across locations, training the network to reconstruct a dropped visit from neighboring visits; that variant would attack the longitudinal variability that most directly inflates pointwise regression slopes.
  • Comparing the p-value-augmented denoiser against an input of total deviation values, which already encode the same normative reference, would test whether the p-value gain comes from semantic side information or simply from the extra input dimension.
  • An ablation with p-values supplied only at masked locations would show whether the network uses them to repair missing data points or to reweight the whole field during encoding.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 3 minor

Summary. The manuscript proposes self-supervised denoising of visual field (VF) data using variational autoencoders and masked autoencoders, with and without categorical per-location p-values as auxiliary inputs. The authors claim that including p-values improves denoising, that masked autoencoders outperform variational autoencoders, and that the proposed approach increases detection of progressing eyes and shortens time to progression. The paper evaluates the methods on a large clinical dataset of 4,232 patients and on simulated VF data with known ground truth, using PLR, MD, and GRI progression criteria.

Significance. If the claims were supported, the work would be clinically useful: denoised VF data that improve the accuracy and earliness of glaucoma progression detection would have immediate translational value for perimetry-based monitoring. The strengths of the paper include a large real-world dataset, a simulation framework with known ground truth, and a publicly available code repository (GitHub link provided). However, the central claim that p-values improve progression detection is contradicted by the paper's own controlled simulation results, and the abstract's headline numbers do not match the results tables. The lack of a clear statement that progression analysis was restricted to the held-out test set further undermines confidence in the clinical results.

major comments (4)
  1. [Table 1, Results] The controlled simulation, which is the only setting with known ground truth, contradicts the central claim that p-values improve detection. Under PLR, the masked autoencoder without p-values detects more progressing eyes than with p-values in every glaucoma progression scenario (slow: 67.49% vs 65.69%; medium: 66.09% vs 64.10%; fast: 65.96% vs 62.63%). Under GRI, the same pattern holds (slow: 77.13% vs 68.95%; medium: 76.00% vs 69.68%; fast: 75.33% vs 68.35%). Only MD shows a slight advantage for p-values, but MD is near ceiling (~99%) for both variants and raw data already achieve ~100% detection, so the MD comparison is uninformative. These results directly undercut the abstract's assertion that 'including a categorical p-value at every visual field location improves the smoothing of visual field data.'
  2. [Abstract, Discussion, Table 2a] The abstract states 'A 4.7% increase in detection of progressing eyes with pointwise linear regression (PLR) was observed,' but no table supports this figure. Table 2a shows MAE+p detects 21.21% versus MAE's 19.62%, a 1.59 percentage-point difference, and the Discussion explicitly says 1.59%. Furthermore, the abstract's claim that p-values predict progression 2.3 months earlier 'when p-values were included' for both masked and variational autoencoders is contradicted by Table 2b: for VAE, conversion time is later with p-values than without for PLR (6.26 vs 6.07 years) and GRI (5.39 vs 4.50 years). The only 0.19-year earlier conversion is for MAE under PLR (7.01 vs 7.20 years). The abstract therefore misrepresents the paper's own numerical results.
  3. [Progression Analysis Methods, Training Details] The manuscript describes a train/validation/test split for model selection, but the Progression Analysis Methods section does not state that only the held-out test set was used to compute the progression statistics in Table 2 and the Kaplan-Meier analyses. The text says 'The VF data at each of the 52 test locations underwent noise reduction and reconstruction using both masked autoencoder and variational autoencoder algorithms,' where 'test locations' appears to refer to the 52 VF spatial locations rather than a held-out set of eyes. If the denoising networks were applied to training eyes when computing clinical progression rates, the small differences in Table 2 could be in-sample artifacts. The authors must clarify whether the progression analysis was restricted to the test set and, if not, re-run the analysis on the test set only.
  4. [Table 2, Kaplan-Meier Analysis] No confidence intervals or p-values are reported for the progression-rate differences in Table 2, and the Kaplan-Meier curves in Figure 4 show overlapping confidence intervals; the text itself acknowledges 'considerable overlap in confidence intervals suggests a potential lack of statistical significance.' Without formal significance testing, a 1.59 percentage-point PLR difference (MAE+p vs MAE) cannot be distinguished from chance variation, especially given that the paper reports no patient-level clustering or multiple-comparison adjustments across the three progression metrics.
minor comments (3)
  1. [Introduction] There are typographical errors in the Introduction, such as 'utilized utilized' and 'outputtedsuch,' and the phrase 'the same was as 15' in the Simulations subsection should read 'the same way as in 15.'
  2. [Figures] The text references to figures are inconsistent: the masked autoencoder is referred to as 'Figure 4' in the Methods but the corresponding figure is labeled Figure 3, and the Kaplan-Meier discussion refers to 'Figure 2' and 'Figure 4' for the same plots. Please align the figure captions and in-text citations.
  3. [Discussion] The Discussion says 'These improvements from masked autoencoders + p-values are visualized in Table 1, depicting results for the PLR and mean deviation metrics,' but Table 1 is the simulation results table, not the clinical results; this is presumably meant to refer to Table 2.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central claim is empirically benchmarked against independent progression criteria, and the self-citations are not load-bearing.

full rationale

The paper's central claim is that adding categorical p-values at each visual field location improves self-supervised denoising and thereby improves glaucoma progression detection. This is tested empirically: four denoisers (VAE and masked autoencoder, each with and without p-values) are trained on visual field data, and their denoised outputs are fed into fixed clinical progression algorithms (PLR, MD, GRI) and compared against raw data and against simulated visual fields with known ground-truth progression. The progression algorithms are external to the training objective, so the evaluation is not defined in terms of the network's fitted parameters. The p-value inputs are a fixed, hand-specified feature derived from total-deviation probability maps rather than a fitted parameter renamed as a prediction. The self-citations present are to the team's prior VAE smoothing work (ref. 17) and to their previously published simulation algorithm (ref. 15); the paper explicitly replicates the VAE on its own dataset rather than importing the result, and the simulation algorithm is only a data generator used to create scenarios with known progression. Neither citation is used to assert a uniqueness theorem or to forbid alternative explanations, and neither defines the claimed p-value benefit. No equation in the paper reduces the reported progression improvement to the model's own inputs by construction. The numerical inconsistencies noted by a skeptical reader (e.g., 4.7% in the abstract vs 1.59% in the Discussion) are reporting discrepancies and do not constitute circular reasoning. The paper is therefore self-contained with respect to its central empirical claim, and no circular step can be exhibited.

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

The paper relies on standard NN training hyperparameters and on domain assumptions about VF p-values and simulation fidelity. No new entities are postulated. The key load-bearing assumption is that progression metrics are computed on data independent of training, which is not explicitly confirmed.

free parameters (3)
  • Number of masked VF locations = 10
    Chosen by hand; no sensitivity analysis; central to MAE denoising behavior.
  • Bottleneck latent size = 16
    Chosen for both VAE and MAE; no ablation.
  • Learning rate = 0.0001
    Standard Adam default; not tuned per model.
assumptions (4)
  • domain assumption Visual field p-values are informative for denoising and can be concatenated as hot-encoded categorical variables without leakage.
    The p-values are derived from the same total deviation data, but the paper treats them as independent auxiliary inputs (Methods, Neural Denoising).
  • domain assumption The simulation algorithm of Rabiolo et al. (ref 15) generates realistic VF noise and progression patterns.
    Simulations are used to validate denoising; if the noise model is wrong, simulation conclusions do not transfer to clinical data.
  • domain assumption PLR, MD, and GRI criteria are valid and sufficient for defining glaucoma progression in both simulated and real data.
    Used as ground truth for progression; no external validation.
  • domain assumption The progression analysis is independent of the training data, despite no explicit statement that only the held-out test set was used.
    The Methods section describes a train/validation/test split but the Progression Analysis Methods section does not confirm test-only evaluation; if violated, clinical results could be in-sample artifacts.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Self-supervised denoising of visual field data improves detection of glaucoma progression." pith.science (2026). https://pith.science/paper/FBTQKHLB

@misc{pith2026241112146,
  author       = {Pith},
  title        = {Pith review of: Self-supervised denoising of visual field data improves detection of glaucoma progression},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FBTQKHLB}},
  note         = {Machine review of arXiv:2411.12146}
}
read the original abstract

Perimetric measurements provide insight into a patient's peripheral vision and day-to-day functioning and are the main outcome measure for identifying progression of visual damage from glaucoma. However, visual field data can be noisy, exhibiting high variance, especially with increasing damage. In this study, we demonstrate the utility of self-supervised deep learning in denoising visual field data from over 4000 patients to enhance its signal-to-noise ratio and its ability to detect true glaucoma progression. We deployed both a variational autoencoder (VAE) and a masked autoencoder to determine which self-supervised model best smooths the visual field data while reconstructing salient features that are less noisy and more predictive of worsening disease. Our results indicate that including a categorical p-value at every visual field location improves the smoothing of visual field data. Masked autoencoders led to cleaner denoised data than previous methods, such as variational autoencoders. A 4.7% increase in detection of progressing eyes with pointwise linear regression (PLR) was observed. The masked and variational autoencoders' smoothed data predicted glaucoma progression 2.3 months earlier when p-values were included compared to when they were not. The faster prediction of time to progression (TTP) and the higher percentage progression detected support our hypothesis that masking out visual field elements during training while including p-values at each location would improve the task of detection of visual field progression. Our study has clinically relevant implications regarding masking when training neural networks to denoise visual field data, resulting in earlier and more accurate detection of glaucoma progression. This denoising model can be integrated into future models for visual field analysis to enhance detection of glaucoma progression.

Figures

Figures reproduced from arXiv: 2411.12146 by the authors.

Figure 1
Figure 1. Depiction of visual field exams from three different patients. The top row shows the right eye (OD) and the bottom row depics the left eye (OS), both depicting both the numerical and grayscale representations of central 24-2 threshold testing. networks to predict final visual field data with a variety of methods and inputs. It is clear that visual field measurements are a cornerstone to better detect, understand, an… view at source ↗
Figure 2
Figure 2. Illustration of the VAE we deployed for this study, where we employed the basic probabilistic encoder-decoder architecture. Through the compression of the visual field data in the latent distribution, the goal is to make it difficult for the probabilistic decoder to reconstruct the noise in the data. both models, we experimented with including a p-value at each location for improved denoising. Variational autoencode… view at source ↗
Figure 3
Figure 3. Depiction of the masked autoencoder constructed. We first zero out 10 random visual field locations, then combine the vectorized visual field locations with the given p-values. Finally, we pass the data through a simple feed forward neural network to reconstruct the pre-masked data. + p-values. We split our dataset into a standard training/validation/testing split and saved the epoch with the lowest validation loss … view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Plot of Kaplan Meier survival curves for the mean deviation and pointwise linear regression metric showing that the masked-autoencoder with p-values outperform the VAE and raw data in survival analysis. Left and right are Kaplan Meier curves for PLR and MD respectively…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

27 extracted references · 20 canonical work pages

  1. [1]

    Reis, A. S. et al. Rates of change in the visual field and optic disc in patients with distinct patterns of glaucomatous optic disc damage. Ophthalmology 119, 294–303 (2012)

    work page 2012

  2. [2]

    Heijl, A. et al. Natural history of open-angle glaucoma. Ophthalmology 116, 2271–2276 (2009)

    work page 2009

  3. [3]

    Nouri-Mahdavi, K. et al. Predictive factors for glaucomatous visual field progression in the advanced glaucoma intervention study. Ophthalmology 111, 1627–1635 (2004). 8/10

    work page 2004

  4. [4]

    Jackson, A. B. et al. Fast progressors in glaucoma: Prevalence based on global and central visual field loss. Ophthalmology 130, 462–468 (2023)

    work page 2023

  5. [5]

    Besharati, S. et al. Detecting fast progressors: Comparing a bayesian longitudinal model to linear regression for detecting structural changes in glaucoma. Am. J. Ophthalmol. 261, 85–94 (2024)

    work page 2024

  6. [6]

    N., Aung, T

    Weinreb, R. N., Aung, T. & Medeiros, F. A. The pathophysiology and treatment of glaucoma: a review. Jama 311, 1901–1911 (2014)

    work page 2014

  7. [7]

    The surgical management of glaucoma: A review

    Lim, R. The surgical management of glaucoma: A review. Clin. & Exp. Ophthalmol. 50, 213–231 (2022)

    work page 2022

  8. [8]

    D., Khawaja, A

    Stein, J. D., Khawaja, A. P. & Weizer, J. S. Glaucoma in adults—screening, diagnosis, and management: a review. Jama 325, 164–174 (2021)

    work page 2021

Show all 27 references
  1. [9]

    Camp, A. S. & Weinreb, R. N. Will perimetry be performed to monitor glaucoma in 2025? Ophthalmology 124, S71–S75 (2017)

    2017

  2. [10]

    H., Choi, H., Lee, S

    Moon, S., Lee, J. H., Choi, H., Lee, S. Y . & Lee, J. Deep learning approaches to predict 10-2 visual field from wide-field swept-source optical coherence tomography en face images in glaucoma. Sci. Reports 12, 21041 (2022)

    2022

  3. [11]

    Prediction of visual field progression with serial optic disc photographs using deep learning

    Mohammadzadeh, V .et al. Prediction of visual field progression with serial optic disc photographs using deep learning. Br. J. Ophthalmol. (2023)

    2023

  4. [12]

    Wu, S. et al. Auxiliary-domain learning for a functional prediction of glaucoma progression. In International Workshop on Ophthalmic Medical Image Analysis , 21–31 (Springer, 2023)

    2023

  5. [13]

    Prediction of visual field progression with baseline and longitudinal structural measurements using deep learning

    Mohammadzadeh, V .et al. Prediction of visual field progression with baseline and longitudinal structural measurements using deep learning. Am. J. Ophthalmol. (2024)

    2024

  6. [14]

    Shuldiner, S. R. et al. Predicting eyes at risk for rapid glaucoma progression based on an initial visual field test using machine learning. PloS one 16, e0249856 (2021)

    2021

  7. [15]

    Rabiolo, A. et al. Quantification of visual field variability in glaucoma: implications for visual field prediction and modeling. Transl. vision science & technology 8, 25–25 (2019)

    2019

  8. [16]

    A., Lemij, H

    Sabouri, S., Pourahmad, S., Vermeer, K. A., Lemij, H. G. & Yousefi, S. Pointwise and region-wise course of visual field loss in patients with glaucoma. Transl. Vis. Sci. & Technol. 11, 20–20 (2022)

    2022

  9. [17]

    Efficacy of smoothing algorithms to enhance detection of visual field progression in glaucoma

    Mohammadzadeh, V .et al. Efficacy of smoothing algorithms to enhance detection of visual field progression in glaucoma. Ophthalmol. Sci. 100423 (2023)

    2023

  10. [18]

    Kingma, D. P. & Welling, M. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013)

    2013 arXiv

  11. [19]

    Zhang, Y ., Liu, W., Chen, Z., Wang, J. & Li, K. On the properties of kullback-leibler divergence between multivariate gaussian distributions. arXiv preprint arXiv:2102.05485 (2021)

    2021 arXiv

  12. [20]

    Agarap, A. F. Deep learning using rectified linear units (relu). arXiv preprint arXiv:1803.08375 (2018)

    2018 arXiv

  13. [21]

    He, K. et al. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 16000–16009 (2022)

    2022

  14. [22]

    Dosovitskiy, A. et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)

    2020 arXiv

  15. [23]

    Kingma, D. P. & Ba, J. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)

    2014 arXiv

  16. [24]

    Sak, H., Senior, A. W. & Beaufays, F. Long short-term memory recurrent neural network architectures for large scale acoustic modeling. Google Interspeech 2014 (2014)

    2014

  17. [25]

    Vaswani, A. et al. Attention is all you need. Adv. neural information processing systems 30 (2017)

    2017

  18. [26]

    & Nash, R

    O’Shea, K. & Nash, R. An introduction to convolutional neural networks. arXiv preprint arXiv:1511.08458 (2015)

    2015 arXiv

  19. [27]

    & Hinton, G

    LeCun, Y ., Bengio, Y . & Hinton, G. Deep learning.nature 521, 436–444 (2015). Acknowledgements We would like to thank the Keck Foundation for their grant to Pepperdine University to support our Data Science program and this research. 9/10 Author contributions statement F.S., ...

    2015

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.