Virtual CT Myelography: A Patch-Based Machine Learning Model to Improve Intraspinal Soft Tissue Visualization on Unenhanced Dual-Energy Lumbar Spine CT
, Engin Dikici 1,*, Sema Candemir 1,†, Daniel J. Boulter 2 and Luciano M. Prevedello 1,2Round 1
Reviewer 1 Report
This paper has adequately described a patch-based machine learning method to help differentiate among major tissue types in the spine, with emphasis on differentiating spinal cord from CSF.
The research topic of this paper is good, and I suggest the authors further improve this paper to meet the other requirements of the publication.
1. The image quality in Fig. 1, 2, 3, and 4 should be improved, a screenshot figure is unacceptable. Fig. 1 (A) and (B) should be adjusted to the appropriate size.
2. The tables and the caption of table should be in the right places with right format, but not put all of them on one page.
3. I think table 3 show that the final ensemble model achieve a quite good prediction result. My only suggestion is the author could show some false classified images to show that “water” and “soft tissue”, “water” and “bone” are hard to distinguish.
4. I am interested in the mentioned output visualization methods and I suggest the author to show the 4 tissue classes in different color, which might be more distinct than 8-bit gray scale visualization.
Author Response
We appreciate the helpful and constructive reviewer comments. All comments have been addressed in bold, and we have made modifications to the manuscript based on this valuable feedback.
Reviewer 1
This paper has adequately described a patch-based machine learning method to help differentiate among major tissue types in the spine, with emphasis on differentiating spinal cord from CSF.
The research topic of this paper is good, and I suggest the authors further improve this paper to meet the other requirements of the publication.
- The image quality in Fig. 1, 2, 3, and 4 should be improved, a screenshot figure is unacceptable. Fig. 1 (A) and (B) should be adjusted to the appropriate size.
Because the intent of Figure 1 is to illustrate the selection of ROIs, particularly those that cover a very small pixel volume, the use of an enlarged, pixelated form of the source images was intentional. However, we have resized the Figure 1 images to a size that would be reasonable for print publication. Figures 2, 3, and 4 have also been similarly resized and appear to show acceptable resolution at their intended print sizes based on our visual assessment. We are willing to continue working with the editorial staff to resubmit additional higher-resolution images should this be deemed necessary.
- The tables and the caption of table should be in the right places with right format, but not put all of them on one page.
All figures, tables, and captions are now in the main text in their appropriate positions, with minor formatting changes to improve readability.
- I think table 3 show that the final ensemble model achieve a quite good prediction result. My only suggestion is the author could show some false classified images to show that “water” and “soft tissue”, “water” and “bone” are hard to distinguish.
The input data are in the form small voxel patches (e.g., 3x3x3x2 arrays of numbers), as shown in the bottom row of Figure 2. In addition to challenges related to displaying 4D data, these patches are also likely too abstract to enable visual qualitative assessment of whether a given patch correctly identifies soft tissue from water. We believe the most intuitive way to visualize examples of true and false classifications is to compare the output images in Figure 6 with the corresponding CT source images. We have created an additional Figure 7, depicting a magnified set of input and output images, with annotations to describe examples of true and false voxel tissue classifications in selected locations.
- I am interested in the mentioned output visualization methods and I suggest the author to show the 4 tissue classes in different color, which might be more distinct than 8-bit gray scale visualization.
The output format is in 8-bit grayscale, but we only used 4 pixel value assignments, so the data represented in each output pixel is effectively capturable in only 2 of the 8 bits. While color depiction can be helpful in situations where more data are present (e.g., different axes of grayscale data), in our case, we have much less data encoded in the output, and therefore, a color depiction would likely be superfluous. However, we acknowledge that it is beneficial to show the output visualization more clearly, and I believe the magnified annotated output image in the new Figure 7 would serve this purpose.
Reviewer 2 Report
The submitted manuscript is devoted to developing a recognition model to distinguish between the spinal cord and cerebrospinal fluid in CT images. The recognition problem lies in a similar mass density, i.e., the spinal cord and cerebrospinal fluid are challenging to separate based on a CT image. The proposed approach based on an ensemble deep learning model is unique and responsibly claims to be a scientific novelty.
Although, there are a few comments concerning the design of the article that may improve the submission:
1. The review in the Introduction might not sufficiently convince a reader about the study's relevance. Thus, it should be updated with the analysis of related works.
2. The Conclusion section should be extended using: 1) numerical results obtained in the paper; 2) limitations of the proposed method; 3) perspectives of the future research. All this information is presented in the article; however, it must also be reflected in the conclusions.
Overall, with regard to the high quality and originality of the presented work, the submitted manuscript is recommended for publication as it is.
Author Response
We appreciate the helpful and constructive reviewer comments. All comments have been addressed in bold, and we have made modifications to the manuscript based on this valuable feedback.
Reviewer 2
The submitted manuscript is devoted to developing a recognition model to distinguish between the spinal cord and cerebrospinal fluid in CT images. The recognition problem lies in a similar mass density, i.e., the spinal cord and cerebrospinal fluid are challenging to separate based on a CT image. The proposed approach based on an ensemble deep learning model is unique and responsibly claims to be a scientific novelty.
Although, there are a few comments concerning the design of the article that may improve the submission:
- The review in the Introduction might not sufficiently convince a reader about the study's relevance. Thus, it should be updated with the analysis of related works.
More details have been added related to the cited prior work in the Introduction to emphasize our study’s relevance to existing literature, particularly our aim of differentiating intraspinal water-type pixels from soft-tissue pixels on CT, in contrast to prior spine CT studies predominantly focused on segmenting bone pixels from non-bone pixels. In addition, text has been added to the Discussion to compare our findings to the existing literature.
- The Conclusion section should be extended using: 1) numerical results obtained in the paper; 2) limitations of the proposed method; 3) perspectives of the future research. All this information is presented in the article; however, it must also be reflected in the conclusions.
The Conclusion section has been updated to include all the requested information. In addition, more numerical results were added to the first paragraph of discussion and briefly summarized in the Conclusions.
Overall, with regard to the high quality and originality of the presented work, the submitted manuscript is recommended for publication as it is.
