Pith. sign in

REVIEW 5 major objections 4 minor 20 references

A System for Accurate Tracking and Video Recordings of Rodent Eye Movements using Convolutional Neural Networks for Biomedical Image Segmentation

T0 review · 5 major / 4 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 This paper claims that two U-Net segmentation models identify rodent pupils and corneal reflections with 99.51% and 98.02% accuracy, making this the first practical deep-learning system for rodent gaze tracking.

desk verdict Useful application of U-Net to rodent eye tracking, but the reported accuracies are training curves, so the central claim of robust performance is unsupported. read the letter →

arxiv 2506.08183 v1 pith:7WQGSJSN submitted 2025-06-09 eess.IV cs.CV

Isha Puri , David Cox This is my paper
classification eess.IVcs.CV
keywords rodenteyetrackingU-Netbiomedicalimagesegmentationpupildetectioncornealreflectioninfraredvideotrackergazeestimationconvolutionalneuralnetwork
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 sets out to replace a fragile, multi-stage image-processing pipeline for rodent eye tracking with a single deep-learning segmentation stage. It claims that two U-Net models, one for the pupil and one for the corneal reflection, reach 99.51% and 98.02% accuracy on the features needed to compute gaze angle. The motivation is that rodents do not reliably saccade to calibration targets, so a non-invasive, self-calibrating video system is essential, and algorithms built for human eyes fail on rodent images with hair, small eyes, and variable eye parameters. If the accuracy claims hold, the system would give neuroscience laboratories a real-time, non-invasive rodent gaze tracker that can be retrained in minutes for new subjects.

What carries the argument

The load-bearing object is the U-Net convolutional architecture, used as two separately trained segmentation models, one for the pupil and one for the corneal reflection, trained end-to-end on manually generated mask ground truth from a few hundred infrared frames. U-Net's contracting path captures context, its symmetric expanding path enables precise localization, and it has no fully connected layers, which lets it learn from small training sets. The gaze-angle computation rests on two geometric assumptions supported by rodent schematic-eye literature: the pupil rotates about the center of the eyeball, and the cornea is nearly spherical, so pupil and corneal reflection locations alone determine gaze angle. Incremental retraining within minutes on a P100 GPU is what makes the system adaptable to variability across rodent subjects.

What would settle it

Re-run the two U-Nets on a held-out set of infrared rodent eye frames from new subjects and compare the segmentations against manual labels; if the pupil and corneal reflection accuracies drop materially below 99.51% and 98.02%, the generalization claim fails. A simpler check is to inspect the training code or data split and determine whether the reported curves come from training data only.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that a standard U-Net, trained on only a few hundred manually labeled infrared images, can segment the pupil and corneal reflection in rodent eye images with high reported accuracy: 99.51% for the pupil model and 98.02% for the corneal reflection model. The authors state that this is the first demonstration of a practical convolutional-neural-network segmentation approach for these features in rodent eye images, and that the two models, working with an established infrared video setup, can determine gaze angle. The corneal reflection task is harder because surrounding hairs create many white spots, and the paper's contribution is showing that a single-stage network handles this variability where the prior Sobel-filter, ellipse-fitting, and Starburst pipeline was sensitive to it.

Load-bearing premise

The load-bearing premise is that the reported accuracies reflect performance on images the models have not seen during training; the paper shows training curves and does not state a separate test set or cross-validation, so if the accuracies were computed on training data, the robustness and state-of-the-art claims are unsupported.

Editorial extensions

If this is right

  • A single U-Net stage can replace the prior radial-transform, Sobel, ellipse-fitting, and Starburst pipeline for rodent eye tracking.
  • Neuroscience laboratories could run real-time, non-invasive gaze recording on head-fixed rodents without subject calibration, with sub-250 ms per-frame latency.
  • The system can adapt quickly to variability in eye size, color, and surrounding hair by retraining on a few hundred newly labeled frames.
  • The same two-model segmentation approach may extend to other animal species whose eyes differ substantially from human eyes.

Reading between the lines

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

  • The reported 99.51% and 98.02% accuracies appear to come from training curves rather than a held-out test set or cross-validation, so the generalization claim is an extrapolation until an independent evaluation is reported.
  • A direct comparison against the previous pipeline's accuracy on the same frames would quantify how much robustness is actually gained, since the old pipeline was not evaluated head-to-head here.
  • The incremental-training claim could be tested by measuring accuracy after retraining on a second subject with a deliberately smaller label budget.
  • Because the gaze-angle computation depends on the near-spherical cornea assumption, the system's angular accuracy is bounded by how well that geometric assumption holds for each rodent strain.
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

5 major / 4 minor

Summary. The paper presents a rodent eye-tracking system that uses two U-Net convolutional neural networks, one to segment the pupil and one to segment the corneal reflection in infrared eye images. The authors report 99.51% accuracy for the pupil model and 98.02% accuracy for the corneal reflection model, and they claim this is the first practical deep-learning segmentation system for these features in rodent eye images. The proposed method is intended to replace a previous multi-stage image-processing pipeline and to enable real-time, non-invasive gaze tracking that can be incrementally retrained for new subjects.

Significance. If properly validated, the approach would be useful for neuroscience and vision-science research, where non-invasive rodent gaze tracking is a practical need. The authors' choice of U-Net is reasonable for small image sets, and the idea of retina/reflection segmentation as a two-task problem is sound. However, the paper's central quantitative claim rests on training-set accuracy, with no test set, no cross-validation, no comparison to the prior pipeline, and no segmentation-quality metric beyond pixelwise accuracy. As presented, the evidence does not support the claimed robustness, state-of-the-art performance, or practical utility, so the contribution is currently more of an early demonstration than a validated method.

major comments (5)
  1. [Section IV, Figure 6] The reported accuracy values (99.51% and 98.02%) are explicitly drawn from training curves, as the caption states "U-Net training curves (Epoch vs Accuracy)". No held-out test set, validation split, cross-validation, or subject-level split is described anywhere in the manuscript. The numbers therefore reflect how well the models fit the training data, not their ability to generalize to unseen images, and they cannot support the headline claims of high accuracy or robustness.
  2. [Section IV] Pixelwise accuracy is not a sufficient metric for this segmentation task because the foreground regions (pupil and corneal reflection) occupy a very small fraction of each image; a degenerate model that always predicts background would achieve high accuracy. The authors should report per-class precision and recall, Dice similarity, or IoU for both classes, and ideally show results on a held-out set with error bars or confidence intervals.
  3. [Section IV] The manuscript claims the new method is superior to the previous pipeline from [16], but it provides no quantitative comparison on the same images or any baseline. The phrases "state of the art" and "highly accurate" require a direct comparison with the previous radial-geometry/Sobel/Starbust pipeline, or at least with a standard segmentation baseline, to be meaningful.
  4. [Section III] Reproducibility is not possible with the information given: the authors state that 245 and 325 labeled images were used for training, but there is no description of how the dataset was split (e.g., by frames vs. by subjects), what loss function and optimizer were used, what hyperparameters were chosen, whether any data augmentation was applied, or what the exact model architecture and input resolution were.
  5. [Section IV] The statement that the model "can be incrementally trained very quickly within minutes on a single P100 Nvidia GPU" is presented as a result, but no timing measurement, training-time experiment, or incremental-learning procedure is described. This claim should be either removed or supported with concrete measurements.
minor comments (4)
  1. [Section I heading] There is a typo in the section heading: "INRODUCTION" should be "INTRODUCTION".
  2. [Figure 6 caption] The caption says "over 99.5% accuracy" and "over 98% accuracy" while the text reports exact values 99.51% and 98.02%; these should be made consistent.
  3. [Figure 4] Figure 4 reproduces the original U-Net figure from Ronneberger et al.; the source should be clearly acknowledged, and permission may be needed for re-use in a journal publication.
  4. [General] The manuscript does not mention approval or compliance with institutional animal care and use regulations, which is typically required for studies involving live rodents.

Circularity Check

1 steps flagged · score 7.0 of 10

Reported 99.51%/98.02% accuracies are training-curve numbers with no held-out evaluation, making the central performance claim a fitted-input result.

  1. fitted input called prediction [Section IV, Results (Figure 6 caption and surrounding text)]
    "We tested our new approach extensively and results show that it can achieve 99.51 % accuracy for the Pupil U -Net model and 98.02% accuracy for the corneal reflection U -Net model (Figure 6). ... Figure 6: U-Net training curves (Epoch vs Accuracy) for rodent eye tracking models: Pupil U-Net Model (top) with over 99.5% accuracy; Corneal Reflection U-Net Model (bottom) with over 98% accuracy"

    The accuracy numbers that constitute the paper's central quantitative claim are explicitly plotted as training curves and are never reported on a held-out test set or cross-validation split. The paper describes only that 'we generated the training masks... using manual labeling for few hundred images' and trained on '245 labeled images for pupil and 325 labeled images for corneal reflection.' Without any evaluation on unseen data, the reported 99.51% and 98.02% accuracies reduce to a measure of how well the models fit the very images used to fit their parameters. This is a fitted input renamed as a performance prediction; it cannot support the 'highly accurate, robust, state-of-the-art' claims for generalization to new rodent subjects.

full rationale

The derivation chain of the paper's central claim is: train two U-Nets on manually labeled rodent eye images, report accuracies, and conclude that the system achieves state-of-the-art accurate and robust tracking. The only quantitative support for this is Figure 6, whose caption explicitly says these are 'training curves (Epoch vs Accuracy).' The text nowhere states that a test set, validation split, or cross-validation was used, and no error bars or confidence intervals are provided. Therefore the headline accuracy figures are training-set accuracies, i.e., measures of the models' fit to the data on which they were trained. Reporting training accuracy as if it were system accuracy is a direct instance of fitted-input-called-prediction: the 'result' is statistically forced by construction. The U-Net architecture itself is cited from an external source (Ronneberger et al.), and the prior eye-tracker hardware is self-cited from [16], but the hardware citation is not what makes the accuracy claim circular; the accuracy claim is circular on its own because the evaluation protocol never separates training from testing. I find no separate self-citation or definitional circularity among the geometric assumptions (spherical cornea, pupil rotation), since those are assumptions drawn from the literature and are not used to define the accuracy numbers. The single load-bearing circular step is the use of training accuracy as the evidence for predictive performance, which warrants a high circularity score because the paper's central 'highly accurate and practical' conclusion rests entirely on that step.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The central claim rests on the validity of the manual labels and the geometric assumptions from prior literature. No new physical entities are introduced. The model weights are fitted to the data, and the reported accuracy is a training-set value, so it does not provide independent evidence for generalization.

free parameters (1)
  • CNN model weights = not reported
    The two U-Net models are trained on 245 and 325 manually labeled images. The reported accuracy is on the training data, so the model performance is a fitted value, not a validated generalization.
assumptions (3)
  • domain assumption Pupil rotates about the center of the eyeball
    Stated in Section II as an assumption supported by rodent eye geometry literature [3,9]. It is necessary to derive gaze angle from pupil and corneal reflection positions.
  • domain assumption Corneal curvature of the rodent is almost an exact sphere
    Also stated in Section II, supported by [3,9]. Needed for the geometric gaze angle calculation.
  • ad hoc to paper Manual labels are accurate ground truth
    The ground truth masks were created by manual labeling, with no inter-annotator agreement or quality control described. The accuracy numbers are measured against these labels.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A System for Accurate Tracking and Video Recordings of Rodent Eye Movements using Convolutional Neural Networks for Biomedical Image Segmentation." pith.science (2026). https://pith.science/paper/7WQGSJSN

@misc{pith2026250608183,
  author       = {Pith},
  title        = {Pith review of: A System for Accurate Tracking and Video Recordings of Rodent Eye Movements using Convolutional Neural Networks for Biomedical Image Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7WQGSJSN}},
  note         = {Machine review of arXiv:2506.08183}
}
read the original abstract

Research in neuroscience and vision science relies heavily on careful measurements of animal subject's gaze direction. Rodents are the most widely studied animal subjects for such research because of their economic advantage and hardiness. Recently, video based eye trackers that use image processing techniques have become a popular option for gaze tracking because they are easy to use and are completely noninvasive. Although significant progress has been made in improving the accuracy and robustness of eye tracking algorithms, unfortunately, almost all of the techniques have focused on human eyes, which does not account for the unique characteristics of the rodent eye images, e.g., variability in eye parameters, abundance of surrounding hair, and their small size. To overcome these unique challenges, this work presents a flexible, robust, and highly accurate model for pupil and corneal reflection identification in rodent gaze determination that can be incrementally trained to account for variability in eye parameters encountered in the field. To the best of our knowledge, this is the first paper that demonstrates a highly accurate and practical biomedical image segmentation based convolutional neural network architecture for pupil and corneal reflection identification in eye images. This new method, in conjunction with our automated infrared videobased eye recording system, offers the state of the art technology in eye tracking for neuroscience and vision science research for rodents.

Figures

Figures reproduced from arXiv: 2506.08183 by the authors.

Figure 2
Figure 2. Previous eye image processing pipeline [16] In the following section, we discuss our new method which deploys Convolutional Neural Networks (CNN) for Biomedical Image Segmentation to overcome these challenges. III. PROPOSED METHOD To overcome the challenges of our problem, this work presents a flexible, robust, and highly accurate model for identification of pupil and corneal reflection. Our new method is a Convolut… view at source ↗
Figure 3
Figure 3. U-Net architecture. The large number of feature channels that compensate for a fully connected neural net layer yield a U-shaped architecture [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 7
Figure 7. A sample video frame processing result, two U-Net models working in conjunction can detect pupil (red) and corneal reflection (green) highly accurately. V. CONCLUSIONS Research in neuroscience and vision science relies heavily on careful measurements of animal subject’s gaze direction. Significant progress in the accuracy and robustness of eye tracking algorithms has so far focused mainly on human eyes, and such alg… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

20 extracted references · 18 canonical work pages

  1. [16]

    Zoccolan, Davide & J Graham, Brett & Cox, David. (2010). A Self-Calibrating, Camera-Based Eye Tracker for the Recording of Rodent Eye Movements. Frontiers in neuroscience

    work page 2010

  2. [1]

    Griffiths S., Scott H., Glover C., Bienemann A., Ghorbel M., Uney J., Brown M., Warburton E., Bashir Z. (2008). Expression of long-term depression underlies visual recognition memory. Neuron58, 186–19410.1016/j.neuron.2008.02.022

    work page arXiv 2008

  3. [2]

    Boraston, Zillah & Blakemore, Sarah -Jayne. (2007). The application of eye -tracking technology in the study of autism. The Journal of physiology. 581. 893-8. 10.1113/jphysiol.2007.133587

    work page Pith review arXiv 2007

  4. [3]

    Hughes A. (1979). A schematic eye for the rat. Vision Res. 19, 569–58810.1016/0042-6989(79)90143-3

    work page arXiv 1979

  5. [4]

    L., Fridman G., Douglas R

    Jacobs A. L., Fridman G., Douglas R. M., Alam N. M., Latham P. E., Prusky G. T., Nirenberg S. (2009). Ruling out and ruling in neural codes. Proc. Natl. Acad. Sci. U.S.A. 106, 5936 – 594110.1073/pnas.0900573106

    work page arXiv 2009

  6. [5]

    F., Lopez -Tellez J

    Lopez -Aranda M. F., Lopez -Tellez J. F., Navarro -Lobato I., Masmudi-Martin M., Gutierrez A., Khan Z. U. (2009). Role of layer 6 of V2 visual cortex in object -recognition memory. Science 325, 87–8910.1126/science.1170869

    work page arXiv 2009

  7. [6]

    F., Mackworth N

    Mackworth J. F., Mackworth N. H. (1959). Eye fixations on changing visual scenes by the television eye -marker. J. Opt. Soc. Am. 48, 439–44510.1364/JOSA.48.000439

    work page arXiv 1959

  8. [7]

    M., Stryker M

    Niell C. M., Stryker M. P. (2008). Highly selective receptive fields in mouse visual cortex. J. Neurosci. 28, 7520 – 753610.1523/JNEUROSCI.0623-08.2008

    work page arXiv 2008

Show all 20 references
  1. [8]

    Remmel R. S. (1984). An inexpensive eye movement monitor using the scleral search coil technique. IEEE Trans. Biomed. Eng. 31, 388–39010.1109/TBME.1984.325352

    1984

  2. [9]

    Remtulla S., Hallett P. E. (1985). A schematic eye for the mouse, and comparisons with the rat. Vision Res. 25, 21 – 3110.1016/0042-6989(85)90076-8

    1985

  3. [10]

    Robinson D. A. (1963). A method of measuring eye movement using a scleral search coil in a magnetic field. IEEE Trans. Biomed. Eng. 10, 137–145

    1963

  4. [11]

    Shackel B. (1967). Eye movement recording by electro - oculography,” in Manual of Psychophysical Methods, eds Venables P. H., Martin I., editors. (Amsterdam: North -Holland; ), 229–334

    1967

  5. [12]

    Stahl J . S. (2004). Eye movements of the murine P/Q calcium channel mutant rocker, and the impact of aging. J. Neurophysiol. 91, 2066–207810.1152/jn.01068.2003

    2004

  6. [13]

    Ronneberger, Olaf. (2017). Invited Talk: U -Net Convolutional Networks for Biomedical Image Segment ation. 3 -3. 10.1007/978 - 3-662-54345-0_3

    2017 doi

  7. [14]

    U-Net: convolutional networks for biomedical image segmentation

    Ronneberger, O., Fischer, P., and Brox, T. U-Net: convolutional networks for biomedical image segmentation. in: N. Navab, J. Hornegger, W.M. Wells, A.F. Frangi (Eds.) Medical Image Computing and Computer -Assisted intervention —MICCAI

  8. [15]

    S., van Alphen A

    Stahl J. S., van Alphen A. M., De Zeeuw C. I. (2000). A comparison of video and magnetic search coil recordings of mouse eye movements. J. Neurosci. Methods 99, 101–11010.1016/S0165- 0270(00)00218-1

    2000 arXiv

  9. [17]

    Xu W., Huang X., Takagaki K., Wu J. Y. (2007). C ompression and reflection of visually evoked cortical waves. Neuron 55, 119 – 129

    2007

  10. [18]

    J., Cox D

    Zoccolan D., Oertelt N., DiCarlo J. J., Cox D. D. (2009). A rodent model for the study of invariant visual object recognition. Proc. Natl. Acad. Sci. U.S.A. 106, 8748 –875310.1073/ pnas. 0811583106

    2009

  11. [19]

    Harezlak K , Kasprowski P , Application of eye tracking in medicine: A survey, research issues and challenges , Comput Med Imaging Graph. 2017

    2017

  12. [2015]

    Springer International, Geneva, Switzerland, 234–241

Pith tools

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