Automated staging of zebrafish embryos with deep learning

  1. Rebecca A Jones
  2. Matthew J Renshaw
  3. David J Barry

Reviewed by

Read the full article See related articles

Listed in

Log in to save this article

Abstract

The zebrafish ( Danio rerio ) is an important biomedical model organism used in many disciplines. The phenomenon of developmental delay in zebrafish embryos has been widely reported as part of a mutant or treatment-induced phenotype. However, the detection and quantification of these delays is often achieved through manual observation, which is both time-consuming and subjective. We present KimmelNet, a deep learning model trained to predict embryo age (hours post fertilisation) from 2D brightfield images. KimmelNet’s predictions agree closely with established staging methods and can detect developmental delays between populations with high confidence using as few as 100 images. Moreover, KimmelNet generalises to previously unseen data, with transfer learning enhancing its performance. With the ability to analyse tens of thousands of standard brightfield microscopy images on a timescale of minutes, we envisage that KimmelNet will be a valuable resource for the developmental biology community. Furthermore, the approach we have used could easily be adapted to generate models for other organisms.

Article activity feed

  1. Version published to 10.26508/lsa.202302351
  2. Review Commons

    Note: This rebuttal was posted by the corresponding author to Review Commons. Content has not been altered except for formatting.

    Learn more at Review Commons

    Reply to the reviewers

    Reviewer #1 (Evidence, reproducibility and clarity):

    In this manuscript the author is presenting a deep-learning model used to predict the development stage of zebrafish embryo. A robust method that can accurately classify a zebrafish into different development stages is highly relevant for many researchers working with zebrafish and hence the importance in developing methods like this is high.

    The manuscript is overall ok. However, more data is needed to convince the reader that the method is robust enough to work with other samples in other labs. This would greatly improve the impact of the publication.

    We agree with the reviewer and have included in our revised manuscripts additional test data that was acquired at a different laboratory to the training data (Figures 5 - 7).

    Page 6.

    • How is the data acquired?

    Images used to do initial model training are the same as those used in a previous study - the details of image acquisition are contained in the relevant publication (doi: 10.12688/wellcomeopenres.18313.1). However, we have now added “Zebrafish Husbandry” and “Live Imaging” for newly-acquired images. We have added a table (Table 1) listing details of all image data used in the study.

    Page 8.
    "This indicates that whileKimmelNet can be used successfully with noisier test data than that on which it was trained,there is an upper limit to how noisy the data can be."

    • This is an obvious statement there will always be an upper limit on noise.

    We agree with the reviewer that this statement is not terribly informative. This section (“KimmelNet’s prediction accuracy is not significantly impacted by moderate levels of additive noise”) has been removed from the revised manuscript in favour of incorporating a section detailing testing of the model on newly-acquired images (“KimmelNet can generalise to previously unseen data”).

    Page 9.

    • Are only wildtype embryos used? How would this work on different mutants. To evaluate the robustness of the method this it would be valuable to test on some mutant line with known developmental difference from the wild type.

    We agree with the reviewer that testing on a mutant line would lend more weight to our findings. For example, the p53-/- zebrafish has a reported, published developmental delay, but we did not have access to that line. However, the developmental delay reported for the p53-/- mutant is virtually indistinguishable from that effected by a temperature change. We therefore focussed our efforts on developmental delay affected by altering incubation temperature only.

    Image data.

    • I would strongly suggest that the author should include a description of the data in the manuscript. A description of how the data is acquired, microscope, different batches, type of embryos used.

    The image data used in the first draft of the manuscript is the same as that used in a previous publication (Jones et al. 2022), which contains all the relevant details the reviewer seeks. However, we have now added the relevant information for the newly-acquired image data.

    "Random160translation in the y-direction was avoided due to the aspect ratio of the images (width>161height) - any artifacts introduced by translation in the x-direction would be removed by the162centre crop layer, but this would not be the case for translation in the y-direction."

    • Could this be solved by using some border method reflection, repetition or fixed value?

    The reviewer is correct that some form of image reflection or repetition could be utilised. However, given the nature of our images, if an embryo is located close to the image boundary, reflection/repetition can result in some odd artefacts, so we minimised the use of such fill methods (also used by the random zoom augmentation layer). We could instead use an arbitrary fixed value, as the reviewer suggested, but finding a value suitable for all images is difficult.

    Page 10.
    Addition of Noise to Image Data

    • This should be added in the training phase. This would probably improve the robustness of the network and also improve the results on the test data.

    We agree with the reviewer and have now added a random Gaussian noise layer for data augmentation purposes during model training (see Figure 1).

    • Supplementary 3 images with high noise. It is worrying that the network is not able to handle the noise in this figure. Looks like the features that is used to distinguish the developmental stage of the embryo is still clearly seen with this high noise level? Retrain the model with noise as an augmentation to improve this.

    As the reviewer suggested, addition of random noise is now incorporated into model training. The new version of the manuscript does not include the same supplemental figures, but instead includes additional testing on newly-acquired data instead.

    Reviewer #1 (Significance):

    The development of methods like this is highly relevant in the zebrafish community. Staging and evaluating the developmental stage for zebrafish is common and is of interest to the broad community. A lot of this work today is done manually, limiting the throughput, and adding human bias.

    The limit of this study is the dataset used for training and evaluation. Firstly, it is not clear about the structure of the data and how it is acquired, different types of fish or imaging setup etc. For a method to be useful to the community it needs to be robust enough to handle different types of fish (transgenic lines). The manuscript would be greatly improved by adding this to the training and evaluation.

    We have now added additional datasets for the purposes of evaluating the model.

    My expertise is image analysis and machine learning for quantification of biological samples, with focus on zebrafish screening.

    Reviewer #2 (Evidence, reproducibility and clarity):

    Summary
    The paper "Automated staging of zebrafish embryos with KimmelNet" by Barry et al., presents a method to automatically stage developmental timepoints of zebrafish embryos based on convolutional neural networks (CNN). The authors show that a CNN trained on ~20k images can determine time post fertilization on test-image sets with an accuracy on the range of a few hours. This technique undoubtedly has the potential to become very useful for any zebrafish researchers interested in developmental timing as it eases analysis and removes potential subjective bias.

    Major comments
    In its current form the paper lacks sufficient graph annotations and method descriptions. This makes it hard in places to judge the validity of the claims. I do believe that the presented method can be useful and is likely valid but to be convincing, the authors need to spend more time expanding the methods, justifying their choices of analysis and clarifying figure annotations.

    We believe that we have addressed the reviewer’s concerns in this revised manuscript, as detailed in response to the specific points below.

    Specific points:

    1. The annotation of the training data is not described and specifically it is unclear how valid the staging of the training data itself is. The authors state in the introduction "the hours post fertilization (hpf) [...] provides only and approximation of the actual developmental stage". It is therefore critical to know how this was accounted for in the annotation of the training data. Since the quality of the training data will ultimately limit the best-case quality of Kimmel Net. The authors need to go into some detail here even though the training data appears to be from another published dataset.

    The reviewer raises a valid point – two individual zebrafish embryos that are x hours post-fertilisation are not necessarily at the same developmental stage. However, we believe it is reasonable to assume that two populations of embryos x hours post-fertilisation are, on average, at the same developmental stage and it is this assumption that forms the basis for our approach. Given the inherent variability the reviewer refers to, we are not suggesting that our model would be particularly accurate for staging individual embryos. However, we are very confident (and we believe the data in the manuscript supports this) that given a population of embryos, our model will provide an accurate rate of development. We have added a paragraph (lines 131-141) to address this point.

    1. Why were "test predictions fit to a straight line through the origin". On the one hand this makes sense (since the slope would indicate the correspondence) but it should be clarified why an intercept was omitted in the fit. After all it is unclear if Kimmel net correctly identifies 0Hpf embryos.

    The reviewer makes a valid point – we do not know what predictions KimmelNet would produce for images of embryos closer to 0 hpf. However, we felt an equation of the form y=mx was a reasonable choice for two reasons. First of all, it matches the form of the Kimmel equation, which, despite its flaws, we are using as a benchmark of sorts – the absence of a y intercept makes comparisons with the Kimmel equation straightforward. Secondly, a “perfect” model would produce a straight line fit with y=x, so the lack of a y intercept seemed a reasonable constraint to impose. We have added some brief text (lines 103-105) to clarify this choice.

    1. The methods do not list how the mean of the absolute error was calculated from 3B/C. I think this should be the mean of the absolute error (not the mean of the error) but in that case the numbers listed in the text appear rather small given the histograms in 3 B/C. A clear statement in the methods would clarify this issue.

    We have now added a “Statistical Analysis” section under Materials & Methods to detail exactly what was used to calculate the values given for error analysis. We have calculated the mean of the error, not the mean of the absolute error, as we wish to illustrate that the mean is close to zero. We have used the standard deviation of the errors to illustrate that there is a significant spread in the error values, as depicted in Figure 3C and D.

    Minor comments

    1. The Y-axis in Figure 2B is simply labeled "Loss" - what is the unit of this loss? HPF? Or HPF^2? This is important for judging the quality of the fit

    We thank the reviewer for drawing our attention to this omission. The loss is hpf2 (mean squared error) and we have updated the plot to reflect this.

    1. Figure 3 B. I would suggest changing the labels of the confidence intervals in the legend. "Inner and outer" is already clear from the figure itself, so labels that state that these are derived from n=100 vs. n=20 test image sized samples would be more useful to the reader

    We thank the reviewer for this suggestion – we have updated the figure legend accordingly.

    Referees cross-commenting

    I concur with comments issued by the other reviewers. I think it will be especially important to address the comments related to testing the method on mutants (Reviewer #1) and training the model in the presence of noise to increase its robustness (Reviewers #1 and #3) as well as addressing the overall annotation/generation of the training data (Reviewers #1 and #2).

    We believe we have now addressed all of these concerns. The model has been retrained with additional data augmentation incorporating random noise, tested on newly-acquired data and we have added tables summarising the details of all image data used in this study.

    I think these points will be critical to make the paper useful by increasing transparency and ensuring reproducibility in other labs with different imaging conditions, strains, mutants, etc.

    Reviewer #2 (Significance):

    Developmental delay is a common occurrence that can be caused by genetic and environmental background effects. It is therefore highly desirable to properly quantify this variable. The work presented here makes an important step in this direction, by allowing to quantify developmental timepoints independent of subjective staging. This speeds up analysis, increases reproducibility and enhances rigor. However, as my comments above indicate, the significance also depends on the ability of potential users to judge the quality of the work. Once those issues have been addressed, I think the work will be of broad interest to the developmental biology community, first and foremost labs utilizing the zebrafish model. However, as the authors state, the presented model architecture could be trained with the data from other species as well.

    Expertise: Zebrafish, quantitative analysis, behavior, neuroscience

    We thank the reviewer for their positive feedback.

    Reviewer #3 (Evidence, reproducibility and clarity):

    Summary:

    Properly staging embryos of zebrafish embryos is important, yet provides challenging since it can depend on many factors, such as temperature, water quality, fish population density, etc. Here, the authors provide a deep-learning-based model called KimmelNet that allows the prediction of the age of zebrafish embryos, using 2D brightfield images. The technique is robust to weak measurement noise and can also be used to identify developmental delays from a very small number of experimental data.

    The code is accessible to the reader, open-source and should be useable by experimentalists without huge effort.

    Major comments:

    I suggest retraining the model and application of the model to additional data for the following reasons:
    • Why did the authors not train for (high) measurement noise and heterogeneous background illumination? Would that not make the model more robust? In principle, creating training should not be considerably harder than before, since the manipulation of the images has been already shown in the manuscript and the authors just need to run it again on the HPC cluster. If there are no technical or administrative constraints (access to the cluster, computational effort, high costs, etc.), the authors should retrain their model.

    We thank the reviewer for this suggestion. As detailed in Figure 1, with a view to making the model more robust, we have now added several more layers of data augmentation, including the addition of random noise, and retrained our model.

    • For Fig. S2 and S3 it is not clear if there is such a strong deviation from the Kimmel equation due to measurement noise or due to the background illumination. The saliency maps appear as if they are mainly affected by the illumination, and maybe less by the noise. Would it be possible to apply the model to a case without artificial noise, but with heterogeneous background illumination to identify what has a bigger impact?

    We thank the reviewer for this suggestion. We have now replaced the “artificial” examples used in the previous version of the manuscript with newly-acquired data (Figure 5), which exhibits different characteristics to that used for training.

    Additionally, the authors need to clarify what exactly they are comparing in this manuscript and rework their interpretation of their findings:
    • When comparing the predictions between KimmelNet and the Kimmel equation, the authors use an equation of the form y=mx. Could the authors please elaborate on why they introduce the constraint of y(0)=0? It might be naturally given by the so-called Kimmel equation, but by looking at Fig 3a, it seems like an equation of the form y=mx+a would be more appropriate and it appears like KimmelNet introduces an offset of around a=2h for 25 Celsius. The authors need to discuss this.

    The main rationale for using an equation of the form y=mx is to be consistent with the Kimmel equation (see lines 103-105). The reviewer is correct that an equation of the form y=mx+c may well produce a better fit to the data, but omitting a y intercept makes comparison with the Kimmel Equation trivial.

    • In lines 5-8 the authors say that developmental stage progression does not only depend on temperature, but also on population density, water quality etc. and they explain that usually not only hpf, but also staging guides based on morphological criteria are used! If that is true, how good is their training data set that only uses hpf and not the other important guides? How did the authors test that these factors have no impact on their training data?

    We have now added a paragraph (lines 131-141) to address this point.

    Since this tool has the potential to have a big impact on zebrafish research, it would be nice to provide some examples of how exactly this could be achieved:
    • Could the authors discuss how exactly their tool is useful to experimentalists? Is it the idea that if an experimentalist wants to investigate an embryo of a particular stage, they apply KimmelNet to images of embryos to identify the stage of the embryo and only then undertake their planned experiment? Is that a realistic undertaking?

    As is evidenced by the errors presented in Figure 3C & D, testing KimmelNet on individual images of embryos may well result in a large error in the predicted hpf. As such, it is not appropriate to use the tool in such a manner. However, to modify the example provided by the reviewer, should an experimentalist have a population of embryos they wished to stage, then yes, KimmelNet would certainly be an appropriate tool for doing so.

    • Would it be possible to provide a tutorial (or even video tutorial if such skills are available in the group of authors) that provides real examples of how to apply the technique? This would make it easier for people without advanced Python/Deep-Learning skills to use the tool, hence improving the impact of KimmelNet.

    A lack of user-friendly interfaces for applying deep learning methods in biology is well-documented – basic knowledge of python and tools like jupyter notebooks are often necessary (https://doi.org/10.1038/s41592-023-01900-4). However, we have endeavoured to make the running of KimmelNet as easy as possible for new users. A jupyter notebook instance can be run on Binder with absolutely no set-up required. To run KimmelNet on their own data, biologists just need to download the Git repo and replace the test images with their own data. We have updated the landing page on the GitHub repo to provide more specific step-by-step instructions for each of these tasks. We will also endeavour to upload our model to the BioImage Model Zoo (https://bioimage.io/#/) to further increase accessibility.

    I am very critical towards equation 1. Please note that I don't think this has any impact on the quality of the technique provided in this manuscript and the significant flaws can already be found in Kimmel 1995 (which is not under review here). That is why I suggest rewriting of this manuscript to not support an over-interpretation of this equation.
    • I do not think that the Kimmel equation is an established term. At least a Google Scholar Search for "Kimmel equation" only gives one result: the preprint of this manuscript.
    • The equation has no mathematical meaning regarding its units (subtracting temperature and a unitless value). I also very rarely see equations with Degrees Celsius and not Kelvin.
    • Additionally, the equation provides a linear relationship between the development time and temperature h(T) and in Kimmel et al, it is shown that this is only true for 25-33 Celsius. Such a linearisation is not very surprising for a small temperature range, but I am not sure how true it is for higher temperature differences. Hence, I feel that it is very bold to give a specific name to such an equation, giving it an importance that it does not deserve.

    We appreciate the reviewer’s concerns and have removed explicit references to “The Kimmel Equation”, without substantively changing the content of the manuscript.

    Minor comments:

    • For the measurement noise cases it would be nice to have some example images of fish with the specific noise levels in Fig S1 and Fig S2.

    We have now removed the “synthetic” additive noise test data, previously depicted in Figures S1-3, in favour of newly-acquired images in Figures 5-7.

    • Could the authors hypothesize why they predict a slower dynamic for 25 Celsius than predicted by the Kimmel equation?

    Referring to Figure 2 in Kimmel et al (1995), it is apparent that the straight lines are by no means perfect fits to the datapoints. In Fig 2A in particular, some datapoints for the 25C data fall well below the line fit. While the published equation suggests a rate of development 80.5% of the rate at 28.5C, according to Fig 2A, an alternative line fit could give a developmental rate as low as 70-75%, which would be in agreement with our data.

    Reviewer #3 (Significance):

    Strengths of the study:

    An easy-to-use method to automatically stage zebrafish embryos and identify differences in the developmental stage is very important for the zebrafish community and the technique in this manuscript definitely novel. The tool is can be used without large effort and the authors suggest that it can also find applications beyond zebrafish embryos. Hence, it is not only interesting to the zebrafish community, but to a broader developmental biology audience.

    Weakness of the study:
    The main weakness of the manuscript is in the training data used for the deep-learning model and the apparent large impact of heterogeneous background illumination. If that is not solved, it is unclear if this technique will find an application in the zebrafish community.

    We believe this weakness has now been addressed by incorporating additional data augmentation measures in the training process and testing the model on newly-acquired data.

    Field of expertise of the reviewer: Image Analysis, Mathematical Modelling, Biological Physics. While I have limited experience in deep learning, I cannot evaluate the specific details of the network architecture. I also have no experience in zebrafish research.

  3. Review Commons

    Note: This preprint has been reviewed by subject experts for Review Commons. Content has not been altered except for formatting.

    Learn more at Review Commons

    Referee #3

    Evidence, reproducibility and clarity

    Summary:

    Properly staging embryos of zebrafish embryos is important, yet provides challenging since it can depend on many factors, such as temperature, water quality, fish population density, etc. Here, the authors provide a deep-learning-based model called KimmelNet that allows the prediction of the age of zebrafish embryos, using 2D brightfield images. The technique is robust to weak measurement noise and can also be used to identify developmental delays from a very small number of experimental data.

    The code is accessible to the reader, open-source and should be useable by experimentalists without huge effort.

    Major comments:

    I suggest retraining the model and application of the model to additional data for the following reasons:

    • Why did the authors not train for (high) measurement noise and heterogeneous background illumination? Would that not make the model more robust? In principle, creating training should not be considerably harder than before, since the manipulation of the images has been already shown in the manuscript and the authors just need to run it again on the HPC cluster. If there are no technical or administrative constraints (access to the cluster, computational effort, high costs, etc.), the authors should retrain their model.
    • For Fig. S2 and S3 it is not clear if there is such a strong deviation from the Kimmel equation due to measurement noise or due to the background illumination. The saliency maps appear as if they are mainly affected by the illumination, and maybe less by the noise. Would it be possible to apply the model to a case without artificial noise, but with heterogeneous background illumination to identify what has a bigger impact?

    Additionally, the authors need to clarify what exactly they are comparing in this manuscript and rework their interpretation of their findings:

    • When comparing the predictions between KimmelNet and the Kimmel equation, the authors use an equation of the form y=mx. Could the authors please elaborate on why they introduce the constraint of y(0)=0? It might be naturally given by the so-called Kimmel equation, but by looking at Fig 3a, it seems like an equation of the form y=mx+a would be more appropriate and it appears like KimmelNet introduces an offset of around a=2h for 25 Celsius. The authors need to discuss this.
    • In lines 5-8 the authors say that developmental stage progression does not only depend on temperature, but also on population density, water quality etc. and they explain that usually not only hpf, but also staging guides based on morphological criteria are used! If that is true, how good is their training data set that only uses hpf and not the other important guides? How did the authors test that these factors have no impact on their training data?

    Since this tool has the potential to have a big impact on zebrafish research, it would be nice to provide some examples of how exactly this could be achieved:

    • Could the authors discuss how exactly their tool is useful to experimentalists? Is it the idea that if an experimentalist wants to investigate an embryo of a particular stage, they apply KimmelNet to images of embryos to identify the stage of the embryo and only then undertake their planned experiment? Is that a realistic undertaking?
    • Would it be possible to provide a tutorial (or even video tutorial if such skills are available in the group of authors) that provides real examples of how to apply the technique? This would make it easier for people without advanced Python/Deep-Learning skills to use the tool, hence improving the impact of KimmelNet.

    I am very critical towards equation 1. Please note that I don't think this has any impact on the quality of the technique provided in this manuscript and the significant flaws can already be found in Kimmel 1995 (which is not under review here). That is why I suggest rewriting of this manuscript to not support an over-interpretation of this equation.

    • I do not think that the Kimmel equation is an established term. At least a Google Scholar Search for "Kimmel equation" only gives one result: the preprint of this manuscript.
    • The equation has no mathematical meaning regarding its units (subtracting temperature and a unitless value). I also very rarely see equations with Degrees Celsius and not Kelvin.
    • Additionally, the equation provides a linear relationship between the development time and temperature h(T) and in Kimmel et al, it is shown that this is only true for 25-33 Celsius. Such a linearisation is not very surprising for a small temperature range, but I am not sure how true it is for higher temperature differences. Hence, I feel that it is very bold to give a specific name to such an equation, giving it an importance that it does not deserve.

    Minor comments:

    • For the measurement noise cases it would be nice to have some example images of fish with the specific noise levels in Fig S1 and Fig S2.
    • Could the authors hypothesize why they predict a slower dynamic for 25 Celsius than predicted by the Kimmel equation?

    Significance

    Strengths of the study:

    An easy-to-use method to automatically stage zebrafish embryos and identify differences in the developmental stage is very important for the zebrafish community and the technique in this manuscript definitely novel. The tool is can be used without large effort and the authors suggest that it can also find applications beyond zebrafish embryos. Hence, it is not only interesting to the zebrafish community, but to a broader developmental biology audience.

    Weakness of the study:

    The main weakness of the manuscript is in the training data used for the deep-learning model and the apparent large impact of heterogeneous background illumination. If that is not solved, it is unclear if this technique will find an application in the zebrafish community.

    Field of expertise of the reviewer:

    Image Analysis, Mathematical Modelling, Biological Physics. While I have limited experience in deep learning, I cannot evaluate the specific details of the network architecture. I also have no experience in zebrafish research.

  4. Review Commons

    Note: This preprint has been reviewed by subject experts for Review Commons. Content has not been altered except for formatting.

    Learn more at Review Commons

    Referee #2

    Evidence, reproducibility and clarity

    Summary

    The paper "Automated staging of zebrafish embryos with KimmelNet" by Barry et al., presents a method to automatically stage developmental timepoints of zebrafish embryos based on convolutional neural networks (CNN). The authors show that a CNN trained on ~20k images can determine time post fertilization on test-image sets with an accuracy on the range of a few hours. This technique undoubtedly has the potential to become very useful for any zebrafish researchers interested in developmental timing as it eases analysis and removes potential subjective bias.

    Major comments

    In its current form the paper lacks sufficient graph annotations and method descriptions. This makes it hard in places to judge the validity of the claims. I do believe that the presented method can be useful and is likely valid but to be convincing, the authors need to spend more time expanding the methods, justifying their choices of analysis and clarifying figure annotations.

    Specific points:

    1. The annotation of the training data is not described and specifically it is unclear how valid the staging of the training data itself is. The authors state in the introduction "the hours post fertilization (hpf) [...] provides only and approximation of the actual developmental stage". It is therefore critical to know how this was accounted for in the annotation of the training data. Since the quality of the training data will ultimately limit the best-case quality of Kimmel Net. The authors need to go into some detail here even though the training data appears to be from another published dataset.
    2. Why were "test predictions fit to a straight line through the origin". On the one hand this makes sense (since the slope would indicate the correspondence) but it should be clarified why an intercept was omitted in the fit. After all it is unclear if Kimmel net correctly identifies 0Hpf embryos.
    3. The methods do not list how the mean of the absolute error was calculated from 3B/C. I think this should be the mean of the absolute error (not the mean of the error) but in that case the numbers listed in the text appear rather small given the histograms in 3 B/C. A clear statement in the methods would clarify this issue.

    Minor comments

    1. The Y-axis in Figure 2B is simply labeled "Loss" - what is the unit of this loss? HPF? Or HPF^2? This is important for judging the quality of the fit
    2. Figure 3 B. I would suggest changing the labels of the confidence intervals in the legend. "Inner and outer" is already clear from the figure itself, so labels that state that these are derived from n=100 vs. n=20 test image sized samples would be more useful to the reader

    Referees cross-commenting

    I concur with comments issued by the other reviewers. I think it will be especially important to address the comments related to testing the method on mutants (Reviewer #1) and training the model in the presence of noise to increase its robustness (Reviewers #1 and #3) as well as addressing the overall annotation/generation of the training data (Reviewers #1 and #2).

    I think these points will be critical to make the paper useful by increasing transparency and ensuring reproducibility in other labs with different imaging conditions, strains, mutants, etc.

    Significance

    Developmental delay is a common occurrence that can be caused by genetic and environmental background effects. It is therefore highly desirable to properly quantify this variable. The work presented here makes an important step in this direction, by allowing to quantify developmental timepoints independent of subjective staging. This speeds up analysis, increases reproducibility and enhances rigor. However, as my comments above indicate, the significance also depends on the ability of potential users to judge the quality of the work. Once those issues have been addressed, I think the work will be of broad interest to the developmental biology community, first and foremost labs utilizing the zebrafish model. However, as the authors state, the presented model architecture could be trained with the data from other species as well.

    Expertise: Zebrafish, quantitative analysis, behavior, neuroscience

  5. Review Commons

    Note: This preprint has been reviewed by subject experts for Review Commons. Content has not been altered except for formatting.

    Learn more at Review Commons

    Referee #1

    Evidence, reproducibility and clarity

    In this manuscript the author is presenting a deep-learning model used to predict the development stage of zebrafish embryo. A robust method that can accurately classify a zebrafish into different development stages is highly relevant for many researchers working with zebrafish and hence the importance in developing methods like this is high.

    The manuscript is overall ok. However, more data is needed to convince the reader that the method is robust enough to work with other samples in other labs. This would greatly improve the impact of the publication.

    Page 6.

    • How is the data acquired?

    Page 8.
    "This indicates that whileKimmelNet can be used successfully with noisier test data than that on which it was trained,there is an upper limit to how noisy the data can be."

    • This is an obvious statement there will always be an upper limit on noise.

    Page 9.

    • Are only wildtype embryos used? How would this work on different mutants. To evaluate the robustness of the method this it would be valuable to test on some mutant line with known developmental difference from the wild type.

    Image data.

    • I would strongly suggest that the author should include a description of the data in the manuscript. A description of how the data is acquired, microscope, different batches, type of embryos used.

    "Random160translation in the y-direction was avoided due to the aspect ratio of the images (width>161height) - any artifacts introduced by translation in the x-direction would be removed by the162centre crop layer, but this would not be the case for translation in the y-direction."

    • Could this be solved by using some border method reflection, repetition or fixed value?

    Page 10.
    Addition of Noise to Image Data

    • This should be added in the training phase. This would probably improve the robustness of the network and also improve the results on the test data.

    • Supplementary 3 images with high noise. It is worrying that the network is not able to handle the noise in this figure. Looks like the features that is used to distinguish the developmental stage of the embryo is still clearly seen with this high noise level? Retrain the model with noise as an augmentation to improve this.

    Significance

    The development of methods like this is highly relevant in the zebrafish community. Staging and evaluating the developmental stage for zebrafish is common and is of interest to the broad community. A lot of this work today is done manually, limiting the throughput, and adding human bias.

    The limit of this study is the dataset used for training and evaluation. Firstly, it is not clear about the structure of the data and how it is acquired, different types of fish or imaging setup etc. For a method to be useful to the community it needs to be robust enough to handle different types of fish (transgenic lines). The manuscript would be greatly improved by adding this to the training and evaluation.

    My expertise is image analysis and machine learning for quantification of biological samples, with focus on zebrafish screening.

  6. Version published to 10.1101/2023.01.13.523922v2 on bioRxiv
  7. Version published to 10.1101/2023.01.13.523922v1 on bioRxiv