An Explainable Classification Method of SPECT Myocardial Perfusion Images in Nuclear Cardiology Using Deep Learning and Grad-CAM

Round 1

Reviewer 1 Report

After the complete study of the above-titled research article, the following comments in bullet points have been jotted down.

·       The research article explores the model as a simple Convolutional network of seven layers i.e., {1-input, 4-Convolutional layer,2-Fully connected layer} solving the image classification problem.

·   This article is supposed to apply the explainable AI concept to every intra-model output using the Grad-CAM technique, but only the model final output is observed with this technique.

·       The image size of the dataset is minimal and blurred to be seen as difficult.

·      Area under the curve (AUC) is described to be shown in the result section but found missing.

·    This research is seemed to be very simple and the inter-paper comparative analysis is not being done, instead the experimental analysis of this research paper technique uses different hyperparameters.

·       The resultant class label by image is missing to have a physical observation of the predicted class.

·       Add the latest article citing the research from deep learning.

·       Standardize all the references as per the template.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Dear authors,

The manuscript is relevant in its results and has a robust sample. However, it is important that you make some adjustments.
A) In figure 3, the graphics can be improved, since they have different sizes and can be placed with a white background. In addition, integrating a measure of dispersion such as the standard deviation would be very useful for the reader and to be able to interpret the results. Delete the horizontal lines and place the scale on the y-axis.
B) About your results of the ROC curves. In step 5, lines 372-374, the procedure is indicated. However, it is important to indicate the assumptions that are required to comply with this type of statistical analysis. In addition, there is no indication of the sensitivity and specificity values ​​(cut-off points) and the discriminative capacity between patients. The AUC, area under the curve, and its confidence intervals (CI 95%) are not indicated for the comparison. each one has an associated estimation error, and it is necessary to report their respective confidence intervals. Diagnostic accuracy from ROC curves generally corresponds to cross-sectional designs. Therefore, it has an associated estimation error, and it is necessary to report their respective confidence intervals.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 3 Report

The authors presented a handcrafted CNN model to classify images of coronary artery disease. The authors gave a detailed description of the dataset and designed the experiments appropriately. The authors applied data augmentation techniques to improve the generalizability of their classification model, and utilized Grad-CAM based color visualization approach to improve the explainability of their model. The overall quality of this paper is good enough in terms of the significance of contents and clarity of the presentation. However, I do have some concerns in regard to the originality of this work. The authors claimed that this work proposed a novel RGB-CNN model. However, I do find other literature that has already discussed similar RGB-CNN models. For instance, [1-4] are some approaches that I found similar to the explainable CNN model proposed in this work. The authors need to make it more clear how their proposed model is different from others and what the model outperforms the others. 

[1] https://arxiv.org/abs/2009.09976

[2] https://ieeexplore.ieee.org/document/8575281

[3] https://link.springer.com/article/10.1007/s44163-021-00015-z

[4] https://www.mdpi.com/2075-4426/11/11/1213

Grad

-CAM-based 32colour visualization approach

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

Dear Authors,

I enjoyed reading this paper.

Still, there are some issues to deal with.

For instance:

• English language and style issues - Grammarly (https://app.grammarly.com) on default settings detected only for the text sample resulting from the concatenation of Title+Abstract+Keywords+Conclusions

8 critical alerts (correctness issues) and
21 more advanced ones, namely:

#Word choice (6),

#Passive voice misuse (5), 

#Punctuation in compound/complex sentences (3),

#Wordy sentences (3),

#Hard to read text (2), 

#and more (2).
This meant a total score of 76 out of a maximum of 100 for this sample above.
Moreover, since you do not appear to be native English speakers, I suggest a total revision of the English language and style for the entire article using Grammarly or another specialized tool;

• The paper mostly meets the structure requirements of the journal, namely:
  Author Information, Abstract, Keywords, Introduction, Materials & Methods, Results, Discussion, Conclusions, etc., as indicated at: https://www.mdpi.com/journal/applsci/instructions 

• You must ensure that all figures have the required resolution (minimum 1000 pixels width/height, or a resolution of 300 dpi or higher according to the Journal’s instructions: https://www.mdpi.com/journal/applsci/instructions  );

• You underlined some items in the Data Prep. step2 (lines 339-364). Is there a hidden meaning of this if compared with the rest of the paper?;

• I think more contributions in journal papers must be cited in this research both in the Introduction and the section dedicated to the interpretation of the results. I also think that just 33 references are not enough for a serious journal paper;

• All digital object identifier (DOI) codes for all journal paper references must be explicitly specified;

• “Pixel size” means the size of pixels (usually measured in nanometers - https://www.justintools.com/unit-conversion/length.php?k1=pixels&k2=nanometers ). I think you meant the picture/image size in pixels or picture/image size (pixels). You should correct all occurrences in the manuscript (e.g. bottom-left of Figure 2-content, Table 3 or lines 356, 442, 445);

• It is not honest to state “The provided dataset” (abstract, line 20) and later (lines 629 and 630) specify “The datasets […] are available from the nuclear medicine physician on reasonable request.” Any research should provide full support for replication of results/reproducibility (https://www.nature.com/articles/533452a );

• You should provide the values for St.Dev., Mean, Median, etc. for all features included in Table 1;

• You must explain why if increasing the image size in pixels (the 5th bolded line in Table 3 vs.  the 8th one), the AUC decreases (0.9458 vs. 0.93). By the way, I think you meant 0.9458 and not 94.58;

• The main contribution of yours should be better highlighted in the summary, end of introduction and conclusions.

Thank you for your contribution!

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

ok . 

Reviewer 4 Report

Dear Authors,

You solved most of the issues pointed out in the 1st round of revisions.

I think your paper is now close to the state of being published.

Sincerely,

D.