Efficient Knowledge Distillation for Brain Tumor Segmentation

Round 1

Reviewer 1 Report

This work proposed a novel approach for Knowledge Distillation for brain tumor segmentation. The proposed improvement tries to integrate channel and space information from the teacher network to the student. It is applied and evaluated on a public dataset of brain MRI images and different networks are compared.

I think that in general the work is valuable and the proposed approach can be of interest. The methodological presentation is quite good, but should be improved in some aspects. I have some comments that the authors should consider to increase the quality of the manuscript.

1.       I found this manuscript similar to the work of Qin et al, 2021 (ref 32 in the manuscript), for both the methodological contribution and the structure of the presentation (sections, types of analyses, results presentation). I suggest the authors to better define and discuss the differences between the two approaches, maybe by also including a quantitative direct comparison, if possible.

2.       In the Experiments section, some details are missing, such as the networks used as teacher and students, that are reported for the first time in the Results section. Image pre-processing should be also better described: what kind of normalization has been adopted? How have the images been cropped? Have the images been resampled? Were the four MRI acquisitions already realigned? Information about training/test split are also missing. The description of the different evaluation and comparison analyses should be reported in this section, rather than in the Results.

3.       The presentation of the results could be improved. For example, figures should represent not only the segmentations, but also the MRI images, in order to better appreciate the subtle differences in the predicted masks. Metric scores should be reported in Tables as mean ± standard deviation.

4.       The Discussion section is currently too poor. Findings should be discussed with respect to the literature, limitations and possible improvements should be reported. The proposed approach is based on a 2D segmentation network. Do you think that a 3D approach may be feasible?

5.       Please check whether all the abbreviations are defined when first cited in the text (e.g. DMI at page 2, line 76 is not defined).

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Interesting work, but need some improvement. In the abstract, it is stated that one of the issues with the current architecture are "computational and storage efficiency issues of segmentation". But this criterion wasn't used for selecting student network. The selection must be better explained. The first paragraph of the introduction should be improved, it is not clear how this research will improve cancer treatment. The second paragraph of the introduction needs more details. The meaning of certain abbreviations is missing like DML. 

For the review of previous work, some detail should be added to better explain the main ideas of the three categories. Also, it should be stated what their main shortcomings are and how this research will improve them.

More details about KD structure as the main contribution of this research should be added. 

It should be written how KL is calculated or referred to additional literature. How hyperparameters are selected, some explanation about choosing values from Table 4 is required. Lacking details about 4 selected network architectures or relevant literature. Also, why these 4 architectures are selected?

In Results, it should be added information about the best scores found in the literature on the same dataset, and to compare these results with the ones from this research.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

I thank the authors for having addressed most of my comments and suggestions. I think that the manuscript has been improved in some aspects. However, there are still some critical issues, maybe due to some errors occurred in the revision process or in the generation of the final pdf.

In the revised version, I didn’t find the section 3 (Methods), where the proposed approach is described. Please, check carefully that all the correct information and details have been preserved during the revision process.

I suggest keeping Experiments and Results sections separated.

Please, check English and try to avoid too verbose or colloquial sentences (e.g. the first sentence in the Discussion is currently too long). Please, try also to avoid the use of cited references as a subject (e.g. “[20] proposed DeepResUNet” should be replaced by “Zhang et al. proposed DeepResUNet [20]”).

Finally, the authors should performed a final text revision to check and correct duplicated information, reported in different sections.

I would prefer having also the final version of the manuscript as a clean pdf, without revision, for a better and easier read.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf