Multimode Fiber Imaging Based on Hydrogel Fiber

JingJing Guo; Lele He; Lili Gui; Mengchao Cao; Xiaosheng Xiao

REVIEW 4 major objections 5 minor 5 references

Multimode Fiber Imaging Based on Hydrogel Fiber

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read The paper demonstrates that speckle patterns from a biocompatible PEGDA hydrogel multimode fiber can be mapped back to recognizable high-resolution handwritten digits by an optimized Pix2Pix image-generation network.

desk verdict Plausible proof-of-concept for hydrogel-fiber imaging, but 'successfully recovered' needs quantitative support before I'd trust it. read the letter →

arxiv 2508.14501 v1 pith:LTUWRJVF submitted 2025-08-20 physics.optics

Lele He , Mengchao Cao , Lili Gui , Jingjing Guo , Xiaosheng Xiao This is my paper · ORCID

classification physics.optics

keywords multimodefiberimaginghydrogelPEGDAspecklereconstructionPix2PixMNISTbiomedicalimage-to-imagetranslation

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to show that a biocompatible PEGDA hydrogel multimode fiber can transmit images well enough for a neural network to recover them. It projects MNIST handwritten digits into the fiber, records the scrambled speckle pattern that emerges, and trains an optimized Pix2Pix image-to-image network to invert that scrambling. The authors report that handwritten digits are reconstructed with recognizable detail, while conventional CNNs such as CVNN, VGG, and ResNet produce blurry or unusable results. The point is to open a route toward flexible, biocompatible imaging fibers for biomedical use, where stiff silica fibers risk damaging tissue.

What carries the argument

The central object is the hydrogel multimode fiber itself: a PEGDA-based soft waveguide whose larger core supports many modes and produces dense speckle. The mechanism that carries the argument is an optimized Pix2Pix generative network—a U-Net generator with seven downsampling and six upsampling layers, skip connections, and a fully convolutional discriminator judging 30×30 local patches—trained end-to-end to map a cropped 256×256 grayscale speckle to the original digit using GAN loss plus L1 loss with weight 200. The preprocessing assumption that the cropped central region retains the light-field information is what lets the network train on manageable, crack-free data.

What would settle it

Train the identical Pix2Pix model on speckles from one hydrogel fiber sample and test it on a second sample prepared the same way, or test it after the first fiber has aged several hours in air; if held-out digit accuracy collapses or the speckle pattern visibly drifts over time, the learned mapping is sample-specific and the claimed imaging capability does not generalize.

Watch

Extended reading notes

Core claim

The central claim is that speckle patterns produced by a PEGDA hydrogel multimode fiber—denser and smaller than those from silica fibers because the hydrogel core guides many more modes—can be mapped back to the original input images by a Pix2Pix generator trained with a combination of adversarial and L1 losses. The paper demonstrates this on 14,000 MNIST image–speckle pairs (80% training, 20% test), preprocessing speckles by cropping to the crack-free central region and linearly increasing brightness. It reports recognizable reconstruction of high-resolution handwritten digits, while CVNN, VGG, and ResNet fail or give only rough shapes, and ImageNet natural images are not yet recoverable.

Load-bearing premise

The method hinges on the belief that the cropped center of a speckle contains most of the image information and that the speckle pattern remains stable over hours of collection, even though the hydrogel fiber is hand-cut, cracked, and losing water.

Editorial extensions

If this is right

Hydrogel multimode fibers become a candidate image-transmission medium for biomedical settings where stiff silica fibers could damage tissue.
Pix2Pix, rather than classification-oriented CNNs, is the better starting architecture for recovering images from dense, highly scrambled speckle.
Handwritten-digit reconstruction at high resolution improves on the coarse 8×8 imaging previously shown with biodegradable polymer fibers.
Natural-image recovery remains out of reach with the current model, marking a clear boundary for future work.
The 14,000-pair dataset and preprocessing recipe provide a baseline for comparing future specialty-fiber imaging systems.

Reading between the lines

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

If the approach transfers to fresh fiber samples, it implies a practical route to implantable or tissue-conformal imaging probes, but that route depends on solving speckle drift and water loss, which the paper does not measure.
The sharp gap between MNIST success and ImageNet failure suggests the Pix2Pix mapping is exploiting the low intra-class variability of digits; a useful next test would be a dataset with intermediate complexity, such as letters or simple shapes, to find where reconstruction degrades.
Because the cropped central region is assumed to carry the information, an ablation that trains on full uncropped speckles versus cropped speckles would isolate whether cropping helps or hurts and directly test the paper's key preprocessing assumption.
The denser speckle from hydrogel fibers may encode more modes, so a physics-informed network that incorporates the fiber's transmission matrix could complement the purely data-driven Pix2Pix approach.

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.

Desk Editor's Note

read the letter

Quick take: this is a legitimate proof-of-concept for imaging through a biocompatible hydrogel multimode fiber, and the claim that Pix2Pix outperforms plain CNNs on this speckle-recovery task is plausible on the evidence shown. But 'successfully recovered' is currently supported only by selected examples; there are no numbers behind it.

What's actually new: the combination of PEGDA hydrogel MMF with a conditional GAN for image reconstruction. The closest prior work I know is citrate fiber at 8x8, and this paper gets recognizable MNIST digits through a hydrogel fiber, which is a real step. The observation that hydrogel speckle is denser and smaller than silica MMF speckle, due to larger core and more modes, is worth noting. The authors also did sensible engineering around the fiber: silica MMF jumper to protect the fiber, silicone tube with deionized water to slow evaporation, and they acknowledge the end-face cracks and manual cutting. The ImageNet failure is honestly reported, which says something about their calibration.

Soft spots, in proportion. The big one is the absence of any quantitative evaluation. No classification accuracy, no SSIM/PSNR, no confusion matrix, no error bars, no comparison against a trivial baseline that always guesses the class prior. With only a handful of selected figures, you cannot distinguish genuine speckle-to-image reconstruction from the GAN exploiting class statistics of MNIST. This is not a fatal flaw—the setup is end-to-end and the L1 loss does constrain pixel-level similarity—but the central empirical claim is not yet established. The stress-test note is right: 'successfully recovered' needs numbers.

Second, the stability of the speckle-transfer function is untested. The fiber loses water (service life extended from 1 hour to ~7 hours), and the end face has cracks. If the speckle drifts during the 14,000-image collection, the trained mapping may not generalize. No stability measurement is reported. Third, the preprocessing choices (cropping to central region, linear brightness enhancement) are justified by the authors' prior work rather than validated on this fiber. That is a mild concern, not circular—no parameter is fitted to force the outcome—but it deserves a check.

Who this is for: people working on specialty fibers for biomedical imaging and on learned reconstruction through scattering media. It is a useful data point, not a breakthrough. I'd send it to peer review because the proof-of-concept is real and the limitations are fixable with a modest amount of additional measurement. My recommendation: major revision requiring held-out quantitative metrics, stability assessment, and a baseline comparison.

Referee Report

4 major / 5 minor

Summary. The paper demonstrates a multimode fiber imaging system using a PEGDA hydrogel fiber as the transmission medium. The authors collect 14,000 MNIST image–speckle pairs, preprocess the speckle images by cropping and brightness enhancement, and train a modified Pix2Pix generator to map speckle patterns back to digit images. They report that Pix2Pix outperforms CVNN, VGG, and ResNet and include qualitative examples in Fig. 4. They also report unsuccessful reconstruction of ImageNet images. The central claim is that high-resolution handwritten images are successfully recovered from hydrogel-fiber speckle patterns.

Significance. If the held-out reconstruction is genuine, this would extend multimode-fiber imaging to a biocompatible hydrogel platform, which is relevant for biomedical applications. The work contributes a new experimental dataset and a task-adapted Pix2Pix architecture. However, the evidence for the central claim is currently qualitative only; no quantitative reconstruction metrics are reported. The paper's own supplementary material discloses several uncontrolled factors (end-face cracks, water loss, manual preparation) that need to be addressed before the empirical claim can be considered established.

major comments (4)

[Section III, Fig. 4] The central claim that images are 'successfully recovered' is supported only by a few selected examples. No quantitative metrics (classification accuracy, SSIM, PSNR, confusion matrix) are reported for the held-out 20% test set. Given that MNIST has only 10 classes and the generator has skip connections, the network could partially rely on class priors and low-frequency speckle statistics rather than faithful speckle-to-image mapping. Please report aggregate test-set metrics and compare against simple baselines (e.g., predicting the unconditional class prior or the mean digit per class).
[Section III, Fig. 4] The comparison among Pix2Pix, CVNN, VGG, and ResNet is based on 'typical' results, with no error bars, no repeated runs, and no statistical test. Since GAN training is stochastic, the observed superiority of Pix2Pix could be due to a favorable run. Please report the distribution of results over multiple training runs, including a quantitative reconstruction metric for each network.
[Supplementary Material, Image Cropping] The decision to crop the speckle to a 256×256 central region is justified by the authors' previous investigations on other fibers, but no validation is provided for the hydrogel fiber. If information in the discarded periphery is necessary for reconstruction, the trained network cannot recover it. Similarly, the 'linear brightness enhancement' is adopted from previous work without evidence that it is appropriate for hydrogel speckle characteristics. Please quantify the information retained by cropping (e.g., compare reconstruction from uncropped vs. cropped speckles) and validate the enhancement step.
[Supplementary Material, hydrogel fiber preparation] The supplementary material discloses that the hydrogel fiber has a service life in air of about 7 hours, that end-face cracks are visible, and that water loss is only slowed by silicone tubing. No speckle stability measurement is reported over the course of collecting the 14,000 image–speckle pairs. If the fiber's transmission properties drift during acquisition due to dehydration or mechanical changes, the training and test sets may not be drawn from the same mapping, undermining generalization. Please report a stability characterization or an acquisition protocol that demonstrates the mapping is stationary.

minor comments (5)

[Conclusion] Typo: 'MINSIT' should be 'MNIST'.
[Supplementary Material, Eqs. (S1)–(S4)] The equations are garbled in the rendered text, making the loss definitions difficult to read. Please provide clean typesetting of the GAN, L1, and discriminator losses.
[Abstract and Section III] The phrase 'high-resolution handwritten images' is not quantified. Specify the input-output image resolution (e.g., 256×256 pixels) and the effective imaging resolution relative to the digit size.
[Fig. 3 caption] The caption uses a full-width colon after OBJ1,2; please standardize punctuation and clarify which objective is used for illumination and which for collection.
[Introduction] Reference [1] is a conference paper on the effects of imperfection and noise; consider citing broader MMF-imaging literature to give context for the Pix2Pix approach and to position the novelty of using hydrogel fibers.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation: the Pix2Pix reconstruction is an empirical supervised-learning claim; the only self-citations are non-load-bearing preprocessing heuristics.

full rationale

The paper's central claim is an empirical supervised-learning result: 14,000 MNIST-speckle pairs from a hydrogel MMF are split 80/20, a Pix2Pix generator is trained with standard GAN and L1 losses (Eqs. S1–S3), and test speckles are mapped to digits (Fig. 4). Nothing in the loss or network architecture defines the test-set outputs; the L1 term is computed against training originals and the adversarial term uses a learned discriminator, so a held-out reconstruction is not forced by construction. The two self-citations in the Supplementary — cropping the central region because "the central area contains most of the original light field information" and linearly increasing speckle brightness "according to our previous study on noise impact" — are preprocessing heuristics inherited from the authors' prior silica-MMF work. They are not load-bearing for the central claim: the sufficiency of cropping and enhancement is empirically falsifiable by the held-out reconstruction, and the paper does not use those citations to prove the reconstruction result. The manuscript does lack quantitative test-set metrics (accuracy, SSIM, PSNR), reports only selected examples, and admits failure on ImageNet; these are rigor and evidence concerns, not circularity. No step of the derivation reduces to its own inputs, so the circularity score is low.

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

The paper is an empirical demonstration with no analytic derivation. The listed free parameters are manually chosen hyperparameters, and the axioms are untested assumptions about fiber stability and information content. The central claim rests on the validity of these assumptions, none of which are quantitatively verified.

free parameters (4)

Pix2Pix L1 loss weight λ = 200
Chosen by hand; balances adversarial and L1 terms; no ablation or cross-validation reported (Supplementary, total generator loss).
Adam learning rate = 0.001
Standard choice, not tuned; affects convergence and final quality (Supplementary).
Adam β1 = 0.5
Standard choice from original Pix2Pix; no sensitivity analysis (Supplementary).
Number of generator downsampling layers = 7 encoding, 6 decoding
Selected to handle the large number of fiber modes; not systematically optimized (Supplementary, Figure S2).

assumptions (4)

domain assumption Cropped central region of the speckle contains most of the original light field information
Stated in the supplementary as 'concluded in our previous investigations' [1], but not re-validated for the hydrogel fiber. If false, the network trains on incomplete data.
domain assumption Speckle patterns are stable over the duration of the experiment
No stability measurement is provided; the fiber suffers from water loss, cracks, and manual handling, so drift is possible.
domain assumption The mapping from input image to speckle is deterministic and learnable from 14,000 paired examples
The entire method assumes a supervised network can learn the inverse mapping and generalize to unseen digits; no information-capacity or generalization-error analysis is given.
domain assumption The hydrogel fiber behaves as a multimode waveguide with sufficient mode diversity
The paper infers larger core diameter and more modes from visual comparison of speckle size, but does not measure core diameter, NA, or number of modes for this fiber.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Multimode Fiber Imaging Based on Hydrogel Fiber." pith.science (2026). https://pith.science/paper/LTUWRJVF

@misc{pith2026250814501,
  author       = {Pith},
  title        = {Pith review of: Multimode Fiber Imaging Based on Hydrogel Fiber},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LTUWRJVF}},
  note         = {Machine review of arXiv:2508.14501}
}

read the original abstract

We demonstrate a multimode fiber imaging technique based on hydrogel fibers, which are suitable for biomedical applications owing to their biocompatibility and environmental friendliness. High-resolution handwritten images are successfully recovered by utilizing a Pix2Pix image generation network.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

5 extracted references · 5 canonical work pages

[1]

Effects of Imperfection and Noise on the Image Reconstruction Through a Multimode Fiber with a Neural Network[C]//2022 Asia Communications and Photonics Conference (ACP)

Cao M, Xiao X. Effects of Imperfection and Noise on the Image Reconstruction Through a Multimode Fiber with a Neural Network[C]//2022 Asia Communications and Photonics Conference (ACP). IEEE, 2022: 77-81

work page 2022
[2]

Soft and stretchable polymeric optical waveguide-based sensors for wearable and biomedical applications[J]

Guo J, Yang C, Dai Q, et al. Soft and stretchable polymeric optical waveguide-based sensors for wearable and biomedical applications[J]. Sensors, 2019, 19(17): 3771

work page 2019
[3]

Soft and stretchable optical waveguide: light delivery and manipulation at complex biointerfaces creating unique windows for on-body sensing[J]

Wu C, Liu X, Ying Y . Soft and stretchable optical waveguide: light delivery and manipulation at complex biointerfaces creating unique windows for on-body sensing[J]. ACS sensors, 2021, 6(4): 1446-1460

work page 2021
[4]

Flexible biodegradable citrate-based polymeric step-index optical fiber[J]

Shan D, Zhang C, Kalaba S, et al. Flexible biodegradable citrate-based polymeric step-index optical fiber[J]. Biomaterials, 2017, 143: 142-148

work page 2017
[5]

speckle-original image

Guo J, Luo Y , Yang C, et al. In situ surface-enhanced Raman scattering sensing with soft and flexible polymer optical fiber probes[J]. Optics Letters, 2018, 43(21): 5443-5446. 8 Supplementary material �� To reduce transmission loss, the preparation conditions of the hydrogel fiber, such as photocuring time, were optimized i...

work page 2018

Pith tools

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

[1] [1]

Effects of Imperfection and Noise on the Image Reconstruction Through a Multimode Fiber with a Neural Network[C]//2022 Asia Communications and Photonics Conference (ACP)

Cao M, Xiao X. Effects of Imperfection and Noise on the Image Reconstruction Through a Multimode Fiber with a Neural Network[C]//2022 Asia Communications and Photonics Conference (ACP). IEEE, 2022: 77-81

work page 2022

[2] [2]

Soft and stretchable polymeric optical waveguide-based sensors for wearable and biomedical applications[J]

Guo J, Yang C, Dai Q, et al. Soft and stretchable polymeric optical waveguide-based sensors for wearable and biomedical applications[J]. Sensors, 2019, 19(17): 3771

work page 2019

[3] [3]

Soft and stretchable optical waveguide: light delivery and manipulation at complex biointerfaces creating unique windows for on-body sensing[J]

Wu C, Liu X, Ying Y . Soft and stretchable optical waveguide: light delivery and manipulation at complex biointerfaces creating unique windows for on-body sensing[J]. ACS sensors, 2021, 6(4): 1446-1460

work page 2021

[4] [4]

Flexible biodegradable citrate-based polymeric step-index optical fiber[J]

Shan D, Zhang C, Kalaba S, et al. Flexible biodegradable citrate-based polymeric step-index optical fiber[J]. Biomaterials, 2017, 143: 142-148

work page 2017

[5] [5]

speckle-original image

Guo J, Luo Y , Yang C, et al. In situ surface-enhanced Raman scattering sensing with soft and flexible polymer optical fiber probes[J]. Optics Letters, 2018, 43(21): 5443-5446. 8 Supplementary material �� To reduce transmission loss, the preparation conditions of the hydrogel fiber, such as photocuring time, were optimized i...

work page 2018