PONE-D-22-24185Segmenting white matter hyperintensities on isotropic three-dimensional Fluid Attenuated Inversion Recovery magnetic resonance images: A comparison of Deep learning tools on a Norwegian national imaging databasePLOS ONE

Dear Dr. Røvang,

Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process.

Ultimately the two Reviewers had conflicting recommendations regarding the decision. Rather than recruiting a third Reviewer, I would like you to carefully address the points raised by both Reviewers.  In addition, please address the following point: - Avoid the use of the word "significant" unless in a statistical context (e.g., line 83)

Please submit your revised manuscript by Nov 20 2022 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at plosone@plos.org. When you're ready to submit your revision, log on to https://www.editorialmanager.com/pone/ and select the 'Submissions Needing Revision' folder to locate your manuscript file.

Please include the following items when submitting your revised manuscript:

• A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.
• A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.
• An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'.

If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter.

If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: https://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory-protocols. Additionally, PLOS ONE offers an option for publishing peer-reviewed Lab Protocol articles, which describe protocols hosted on protocols.io. Read more information on sharing protocols at https://plos.org/protocols?utm_medium=editorial-email&utm_source=authorletters&utm_campaign=protocols.

We look forward to receiving your revised manuscript.

Kind regards,

Niels Bergsland

Academic Editor

PLOS ONE

Journal Requirements:

When submitting your revision, we need you to address these additional requirements.

1. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at 

https://journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_main_body.pdf and 

https://journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatting_sample_title_authors_affiliations.pdf.

2. In your Data Availability statement, you have not specified where the minimal data set underlying the results described in your manuscript can be found. PLOS defines a study's minimal data set as the underlying data used to reach the conclusions drawn in the manuscript and any additional data required to replicate the reported study findings in their entirety. All PLOS journals require that the minimal data set be made fully available. For more information about our data policy, please see http://journals.plos.org/plosone/s/data-availability.

Upon re-submitting your revised manuscript, please upload your study’s minimal underlying data set as either Supporting Information files or to a stable, public repository and include the relevant URLs, DOIs, or accession numbers within your revised cover letter. For a list of acceptable repositories, please see http://journals.plos.org/plosone/s/data-availability#loc-recommended-repositories. Any potentially identifying patient information must be fully anonymized.

Important: If there are ethical or legal restrictions to sharing your data publicly, please explain these restrictions in detail. Please see our guidelines for more information on what we consider unacceptable restrictions to publicly sharing data: http://journals.plos.org/plosone/s/data-availability#loc-unacceptable-data-access-restrictions. Note that it is not acceptable for the authors to be the sole named individuals responsible for ensuring data access.

We will update your Data Availability statement to reflect the information you provide in your cover letter.

3. Please include a separate caption for each figure in your manuscript.

[Note: HTML markup is below. Please do not edit.]

Reviewers' comments:

Reviewer's Responses to Questions

Comments to the Author

1. Is the manuscript technically sound, and do the data support the conclusions?

The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented.

Reviewer #1: Yes

Reviewer #2: No

**********

2. Has the statistical analysis been performed appropriately and rigorously?

Reviewer #1: Yes

Reviewer #2: No

**********

3. Have the authors made all data underlying the findings in their manuscript fully available?

The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified.

Reviewer #1: Yes

Reviewer #2: No

**********

4. Is the manuscript presented in an intelligible fashion and written in standard English?

PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here.

Reviewer #1: Yes

Reviewer #2: No

**********

5. Review Comments to the Author

Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters)

Reviewer #1: The authors investigated the use of deep learning approaches to segment 3D FLAIR data. They collected and used data from a large multisite dementia initiative to test two 3D deep models for segmenting WMH (one trained internally and one without any further training) compared to a Bayesian deep network. The models were trained on a subset of 642 participants and tested on a subset as well as an external dataset. Some subset of data was manually segmented to evaluate and train the model. They found that the 3D nnU-net performed best outperforming other models with some caveats. This is an interesting analysis, however there are some issues that should be addressed:

1. In the 2.6.1 section, many of these decisions are described but not critically reasoned. It would be helpful to justify the reasoning behind various parameter choices in the manuscript (even if they are just based on prior experience and work).

2. The performance on the external dataset is quite low in general. The authors should comment on whether intensity normalization approaches may help generalize these types of approaches.

3. The authors state: “However, since the models tested here were not trained on the MICCAI dataset, a direct comparison of performance may be misleading.” The MICCAI is an open challenge, did the authors consider submitting results to identify how well they performed in that dataset? Otherwise, if this is not important then these types of statements should be removed.

4. The authors state: “Our results do, however, suggest that T1-weighted images may not be needed for WMH segmentation.” Why is this important especially given that a T1-weighted image is almost always available if a FLAIR is available?

5. The conclusions of the manuscript leave a lot to be desired. While the initial motivation for the paper seems relevant, the conclusions are quite flat. For instance, they state that models can be trained on 3D imaging data, it is unclear why this is important to establish. This seems like something that should be feasible. They also state that a T1-weighted image would not be necessary, but this (see previous comment) is also not that important given that they are often available (unless there is motion).

6. The authors also should revisit their hypotheses from the introduction explicitly.

7. It is unclear how many manual segmentations were completed. How many slices were manually segmented?

8. In the introduction, the authors should include a paragraph on why it is important to develop 3D FLAIR segmentation algorithms. What challenges exist in this space? What has been done previously? Are there any tools that have been previously developed for 3D FLAIR? If so, why are they not used here? Or why do they not work?

Minor

1. The figure resolutions on all images are quite poor. They should all be redone to higher resolution images.

2. Tables 3 and 5 show the same results as figures 2 and 3. Just one set can be presented.

3. Some results are presented multiple times in the text, tables, and figures. For instance, for the internal test – the authors state: “For the internal test set, average (+/- std.dev) DSC scores were 0.70 (+/- 0.13), 0.78 (+/-0.10), and 0.63 (+/- 0.15), for the 2.5D U-Net, 3D nnU-Net and Deep Bayesian models, respectively.” They then present this in table 5 and figure 3. It would be better to show it, for instance, in one table and then just state the finding: “3D nnU-Net performed best across all metrics in the internal data (see table 5).”

Reviewer #2: The authors present an evaluation of 3 models/algorithms for segmenting white matter hyper-intensities (WMH) on 3D T2 FLAIR MRI images. One of the models has been implemented and trained by the authors. The second model is the default nnUnet implementation, and the third one is a deep Bayesian model. Two of these models are trained on the Norwegian Disease Dementia Initiation (DDI) dataset, and tested on a subset of the data associated with this study, as well as on an additional external dataset. The main difference between these two networks and the Bayesian one is that the later was trained on combinations of T2-FLAIR and T1 acquisitions.

The main outcomes are: a demonstration that segmenting WMH on 3D T2-FLAIR acquisitions is feasible, and that the nnUnet gives the best evaluation metrics.

English language skills are globally satisfactory, but the paper structure is sometimes difficult to follow. As an example: there is a pre-processing section, 2.5, but data pre-processing (related with the nnUnet) is not detailled here, rather in a later section, 2.6.2. Inconsistencies across section 2.7 and section 2.8 and the results create additional confusion. While test metrics and statistics are presented in section 2.7, presentation of results start with a not introduced section about WMH volume distribution of the data. In addition, TP, FN and FP are used as a performance indicators, but they are not presented in sections 2.7 and 2.8.

In the introduction, the authors assess that so far WMH segmentations have mostly been attempted on 2D FLAIR acquisitions. Could you please provide citations (line 102, line 111 again) ? Nevertheless, one cited paper (Forooshani et al., 2022), is not discussing the use of 2D images. Instead, a 3D Bayesian convolutional neural network is proposed. A confusion between T2-FLAIR and 2D FLAIR may have occurred ?

Moreover, the two articles presented as state-of-the art solutions are citation 5 (citation number 7) and citation 6 (citation number 10): perhaps there is more out there in the literature ? In addition, the Bayesian model cited as SOTA solution requires the use of both T1 and T2 acquisitions, while the objective of the article is to perform segmentation on T2-FLAIR, which creates some confusion.

Some parameters in the equations, as in equation (1), are not explained by the authors (alpha and beta in this case).

The level of details is unbalanced: on one side there is an extensive description of Python libraries for splitting data into train/val subsets, on the other the N4 implementation or DICOM conversion tools are not detailed at all. The latter appears to me as of higher value for reproducibility purpose. Also, details are provided concerning the DDI study, that are of interest, but not of the highest value regarding the work here presented. The reader is being referred to another article to get this valuable information.

Inconsistencies arise between text and tables. Regarding the DDI study, 5 national sites are mentioned, whereas 13 institutions are presented in the corresponding table. The origin or meaning of “institutions” should be explained, as it is difficult for the reader to associate them with the written details. Lines 387-389: three intervals are associated with 4 values, another example of inconsistency.

While the fist model is presented as the in-house model at the beginning of the article, the nnUnet implementation is also referred as an in-house solution in later sections. This creates confusion: the nnUnet is not a solution designed and implemented by the authors (see line 410 just as an example).

The methods are too different to be compared. In fact, they differ in terms of input needs (1 vs 2 contrasts), pre-processing, training and validation splitting (what would have other folds of the nnUnet cross-validation procedure generated as results ?) and prediction (test time augmentation in the case of nnUnet). This level of differences makes it hard to appreciate the cause of different performances. Would some technical choices made in the nnUnet model help the in-house model to learn better on the training set ? I suppose that having at least the same pre-processing and prediction steps would be necessary for sake of comparison.

An additional concern is associated with the use of the results obtained during the MICCAI WMH segmentation challenge as scores to which the proposed algorithms shall be compared. Nevertheless, comparing results obtained on different data in a single table appears as unsuitable to establish relevant conclusions.

Lines 423-428 are unclear. Is the objective of the work to provide a clinically relevant solution, or to obtain the best performance scores ?

In conclusion, the fact that nnUnet performs the segmentation of WMH on 3D T2-FLAIR acquisitions is an interesting result. The comparison with the two other models does not appear relevant because of the level of differences in implementation, inputs requirements and results. The aspect concerning work performed on 2D T2-FLAIR acquisition should be clarified, as 3D acquisitions are a standard.

**********

6. PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files.

If you choose “no”, your identity will remain anonymous but your review may still be made public.

Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy.

Reviewer #1: No

Reviewer #2: No

**********

[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.]

While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, https://pacev2.apexcovantage.com/. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at figures@plos.org. Please note that Supporting Information files do not need this step.

PONE-D-22-24185R1Segmenting White Matter Hyperintensities on Isotropic three-dimensional Fluid Attenuated Inversion Recovery Magnetic Resonance Images: Assessing Deep Learning Tools on a Norwegian Imaging DatabasePLOS ONE

Dear Dr. Røvang,

Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process.

As was the case with the initial submission, the two Reviewers have again provided me with conflicting recommendations. It appears to me that there is still a considerable amount of work that can be done to improve your manuscript. Please carefully consider each of the concerns raised by Reviewer 2.

Please submit your revised manuscript by Mar 12 2023 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at plosone@plos.org. When you're ready to submit your revision, log on to https://www.editorialmanager.com/pone/ and select the 'Submissions Needing Revision' folder to locate your manuscript file.

Please include the following items when submitting your revised manuscript:

• A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.
• A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.
• An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'.

If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter.

If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: https://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory-protocols. Additionally, PLOS ONE offers an option for publishing peer-reviewed Lab Protocol articles, which describe protocols hosted on protocols.io. Read more information on sharing protocols at https://plos.org/protocols?utm_medium=editorial-email&utm_source=authorletters&utm_campaign=protocols.

We look forward to receiving your revised manuscript.

Kind regards,

Niels Bergsland

Academic Editor

PLOS ONE

[Note: HTML markup is below. Please do not edit.]

Reviewers' comments:

Reviewer's Responses to Questions

Comments to the Author

1. If the authors have adequately addressed your comments raised in a previous round of review and you feel that this manuscript is now acceptable for publication, you may indicate that here to bypass the “Comments to the Author” section, enter your conflict of interest statement in the “Confidential to Editor” section, and submit your "Accept" recommendation.

Reviewer #1: All comments have been addressed

Reviewer #2: (No Response)

**********

2. Is the manuscript technically sound, and do the data support the conclusions?

The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented.

Reviewer #1: Yes

Reviewer #2: Partly

**********

3. Has the statistical analysis been performed appropriately and rigorously?

Reviewer #1: Yes

Reviewer #2: Yes

**********

4. Have the authors made all data underlying the findings in their manuscript fully available?

The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified.

Reviewer #1: No

Reviewer #2: No

**********

5. Is the manuscript presented in an intelligible fashion and written in standard English?

PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here.

Reviewer #1: Yes

Reviewer #2: Yes

**********

6. Review Comments to the Author

Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters)

Reviewer #1: (No Response)

Reviewer #2: I would like to recognize and thank the authors regarding the work performed in order to address some of the concerns raised during this first exchange. Notably, precision concerning the different kinds of acquisitions, what was performed in-house and not, the concision effort on some parts of the article and the development of aspects that required additional explanations are very welcome.

Nevertheless, despite these positive improvements, this new version raises a lot of new interrogations, that I would consider at least as important as the ones discussed during the first round.

The objective of the article, from my understanding, is to report work performed in order to provide an automated WMH segmentation from 3D FLAIR images to clinicians. Work has been done and is available in the literature on this topic, but presents differences with the one reported here. Deep learning-based methods handling this segmentation task have been proposed, but based on 2 images, 3D-T1 (2D T1 is still mentioned instead of 3D in the introduction) and 2D FLAIR. Other methods use 3D FLAIR, but are not based on deep learning methods.

I disagree with the authors regarding their response to reviewer 1 (R1.8), as the main objective of this work being to segment WMH from 3D FLAIR data, the other methods related to this application should be compared to the proposed models. Indeed, despite the fact that DL-based methods have been shown to provide increased performance metrics on a number of applications compared to previous works, the “superiority” of the DL-based models needs to be proven against gaussian-mixture models (GMM) or the method proposed in Zhong et al, 2014 for this particular case. Moreover, the authors mention the use of GMM as part of their semi-automated annotation procedure, which I understand as the authors having effectively used this method, which would have then been easily compared. This comparison would have been, in my opinion, at least as meaningful as the comparison with a 2 input images model, from which we could have expected the lower performance metrics, due to the absence of this bayesian model retraining.

The medical images processing community is very familiar with the Unet architecture and the nnUNet framework, but not with the models presented here as state of the art (Hypermapp3r) or the 2.5-D Unet. In particular, the selection of the 2.5-D model should be more motivated, and the choices made to generate it really detailed in the publication, as the reference provided to the readers for explanations is the main author’s master thesis report.

In the results and discussion sections, the nnUNet model is presented as the one providing the overall best performance metrics. From the clinical point of view, is the coverage of the large lesions of the highest importance, or the capability of a model to actually detect lesions, even the smaller ones ? If detecting lesions has a significant meaning for clinicians, is the selection of the nnUNet model still relevant ? For an example of lesions detection evaluation, an interesting reference could be Commowick et al., 2018 (doi: 10.1038/s41598-018-31911-7)

The Hypermapp3r model is presented as having some appreciable advantages. The lower number of parameters, compared to the 2 other models, would be an advantage regarding model’s deployment, and this bayesian model can provide uncertainty related information. But the perspective of retraining / fine-tuning this model or adapting a model with these properties to the 3D-FLAIR WMH application is not envisaged.

The authors did not provide information about the reason of the population description changes (number of subjects modified from 642 to 441, that influences all the related subsets).

Overall, despite interesting aspects presented in this paper, I understand it as a compilation of valuable work, that unfortunately lacks investigations and does not allow to construct and respond to the story that could be attached to the main objective.

**********

7. PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files.

If you choose “no”, your identity will remain anonymous but your review may still be made public.

Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy.

Reviewer #1: No

Reviewer #2: No

**********

[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.]

While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, https://pacev2.apexcovantage.com/. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at figures@plos.org. Please note that Supporting Information files do not need this step.

Segmenting White Matter Hyperintensities on Isotropic three-dimensional Fluid Attenuated Inversion Recovery Magnetic Resonance Images: Assessing Deep Learning Tools on a Norwegian Imaging Database

PONE-D-22-24185R2

Dear Dr. Røvang,

We’re pleased to inform you that your manuscript has been judged scientifically suitable for publication and will be formally accepted for publication once it meets all outstanding technical requirements.

Within one week, you’ll receive an e-mail detailing the required amendments. When these have been addressed, you’ll receive a formal acceptance letter and your manuscript will be scheduled for publication.

An invoice for payment will follow shortly after the formal acceptance. To ensure an efficient process, please log into Editorial Manager at http://www.editorialmanager.com/pone/, click the 'Update My Information' link at the top of the page, and double check that your user information is up-to-date. If you have any billing related questions, please contact our Author Billing department directly at authorbilling@plos.org.

If your institution or institutions have a press office, please notify them about your upcoming paper to help maximize its impact. If they’ll be preparing press materials, please inform our press team as soon as possible -- no later than 48 hours after receiving the formal acceptance. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact onepress@plos.org.

Kind regards,

Niels Bergsland

Academic Editor

PLOS ONE

Additional Editor Comments (optional):

Reviewers' comments: