Repurposing the Public BraTS Dataset for Postoperative Brain Tumour Treatment Response Monitoring
, Claes Nøhr Ladefoged 4,5, Vibeke Andrée Larsen 1, Flemming Littrup Andersen 2,4, Michael Bachmann Nielsen 1,2, Hans Skovgaard Poulsen 3, Jonathan Frederik Carlsen 1,2,† and Adam Espe Hansen 1,2,3,†Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsDL segmentation strategies trained on pre-operative GBM images are suboptimal for segmenting post-operative images, that may contain peculiar features like resection cavities, tissue scarring and hematomas. The authors of this manuscript propose to train a DL segmentation strategy for post-operative GBM images on the pre-operative BraTS 2021 dataset with ground truth labels properly “adapted” to post-operative tumours. Specifically, the 3-label BraTS ground truth were automatically converted to 2-label segmentations (CE and NE), according to rules suitable for both pre-operative and post-operative images. Mainly, the necrosis part and cystic regions were eliminated from the segmentation.
The manuscript is clearly written. However, I think there are not enough elements for a publication. To support the statements made, other evaluations should be performed.
In the following my main concerns.
1) Automatic conversion from 3-labels to 2-labels on BraTS preoperative images.
- BraTS 3-label segmentations were automatically converted into 2-label segmentations by an 8 steps algorithm. At the first step, the whole tumour area (CE+NE) is obtained with HD-GLIO. HD-GLIO itself provide estimates of CE and NE, called here CE_HD-GLIO and NE-HD-GLIO. In the example reported in Figure 1, BraTS 3-label segmentation and HD-GLIO whole tumor segmentation are nearly coherent. How were incoherencies between HD-GLIO and BraTS managed?
- It seems that everything is conceived for circumscribed tumours. What about tumours with other morphologies? Which was the quality of the 2-labels provided by the automatic conversion?
- The conversion has been visually validated on 1/10 of the converted studies. I think the validation needs to be done in more depth.
2) Validity of the proposed post-operative GBM segmentation strategy.
- I am personally not convinced of the proposed idea. An UNet can’t properly manage post-operative image features without having never seen them. The state of the art strategy, HD-GLIO, was indeed trained on a dataset containing both pre-operative and post-operative images.
- The authors say that they could not add post-operative data to the training dataset because they did not have them. However, they have just published a work (reference 13) in which they created a post-operative dataset of 66 studies with manual segmentations to test HD-GLIO performance of post-operative images. I suggest adding it to the training dataset. At this point the idea of augmenting a small post-operative dataset with adapted pre-operative data would make more sense.
- In discussion the authors say that that they obtained “tumour segmentation performance comparable to that of a state of the art algorithm for volumes exceeding 1cm3”. The median Dice was 0.75 for CE and 0.68 for NE; while HD-GLIO obtained 0.79 for CE and 0.71 for NE. Were Dice distributions significantly comparable?
- Why do you think HD-GLIO obtained good results on CE < 1cm3?
Minor points.
- The ability to exclude resection cavities from post-operative images was assessed basing on an automatic script which segmented surgical cavities on post-operative images. Was the cavity segmentation algorithm applied to the entire brain or only to the areas segmented by HD-GLIO? Were the segmentations of these cavities manually refined before quantifying the errors?
- On the preoperative test set, HD-GLIO performs worse than the proposed strategy, on both CE and NE. Please comment on this point. Is the definition of CE and NE assumed by HD-GLIO coherent with the one assumed here?
- In the discussion first sentence, the authors say that “In this study, we demonstrated that using the BraTS dataset to train a DL algorithm for segmentation of brain tumours after surgery will result in substantial and frequent mislabelling of resection cavities.” Where has this point been demonstrated?
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThis study uses an automated conversion of the BraTS annotation protocol to a 2-label system more appropriate the postoperative MRIs. The results show an improvement in accurate resection cavity exclusion But, I still have some reservations.
1. The reliance on the Lumiere dataset may limit the generalizability of the findings.
2. Several references are not well cited.
3. I suggest to include a flowchart in the method section for better understanding by readers.
4. The author's subjective evaluation was conducted on every 10th case, revealing three exceptions. It is recommended to perform a more comprehensive subjective evaluation to ensure thorough analysis.
5. 'Accessed on December 30', year is missing.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for AuthorsComments for author File: Comments.pdf
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsThe inclusion in the appendix of a schematic representation of the steps of the conversion of the BraTS 2021 three-label segmentation masks to the two-label segmentation improved the methods clarity.
Otherwise, the manuscript has not been modified respect to the original version. The explanations included in the reply letter to the reviewer may be included in the discussion of the manuscript.
Author Response
Please see the attachment. Thank you.
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for AuthorsAfter reviewing the revised manuscript, I regret to inform you that the requested changes have not been incorporated into the text. While the responses to the reviewer's report were provided, there were no corresponding modifications made in the manuscript itself.
Therefore, I am returning the manuscript for further revision. Please ensure that all the suggested changes are thoroughly addressed and reflected in the manuscript as previously outlined in my review.
Thank you for your attention to this matter.
Author Response
Please see the attachment. Thank you.
Author Response File: Author Response.pdf
Round 3
Reviewer 1 Report
Comments and Suggestions for Authorsok
Reviewer 3 Report
Comments and Suggestions for AuthorsNow manuscript improved a lot. Recommend to publication.
