Efficient Training on Alzheimer’s Disease Diagnosis with Learnable Weighted Pooling for 3D PET Brain Image Classification

Round 1

Reviewer 1 Report

The paper is well-written. However, the author should do the following minor corrections

1. The contributions of the paper should be listed pointwise in the introduction section.
2. The proposed architecture presented in Section 2.2 should be elaborated in a logical sequence.
3. How the performance scores of different baseline approaches such as 3D-CNN [14], 2D-CNN [17], Jo et al. [15], and ARP [21] presented in Table 4 are obtained? Did you perform the experiment on (ADNI) database? Please explain. 
4. It is suggested to compare the proposed approach with research works of the years 2021, and 2022.

Author Response

We thank the reviewer for the thoughtful comments and suggestions. We were able to address all the comments and revise the manuscript accordingly. Below are the point-by-point responses. Please feel free to let us know if anything we can further provide or clarify. We hope with the revision, the manuscript is now suitable to be published in Electronics.

1. The contributions of the paper should be listed pointwise in the introduction section.

Response: We thank the reviewer for the comment. We added our contributions as pullet points in the Introduction (Lines 76-85).

2. The proposed architecture presented in Section 2.2 should be elaborated in a logical sequence.

Response: We have rewritten section 2.2. accordingly. We hope the current section is more logical.

3. How the performance scores of different baseline approaches such as 3D-CNN [14], 2D-CNN [17], Jo et al. [15], and ARP [21] presented in Table 4 are obtained? Did you perform the experiment on (ADNI) database? Please explain. 

Response: We performed the experiments on the same dataset obtained from ADNI.  To ensure fair comparisons, we conducted the 5-fold cross-validation across all the models to determine the performance.

4. It is suggested to compare the proposed approach with research works of the years 2021 and 2022.

Response: We compared a recent research work of the year 2022 (reference #19) as the baseline in our paper and add the results in Table 4.

Reviewer 2 Report

In this paper a novel 3D-2D fusion CNN based detection method is proposed to detect Alzheimer Diseases. A learnable weighted pooling method is proposed to covert the 3D PET Brain images into 2D ones to reduce the image processing time Moreover, a dual-attention mechanism that includes self attention and channel attention modules is also involved to improve the local information capture and hence improve the detection results. And their experiments demonstrate that the proposed framework outperforms other state-of-the arts in both training and detection performance. It is interesting to see a fast and high performance method to detect Alzheimer diseases.

I just have two minor questions:

1. It is better to use the same symbol I_x for Figure 2 and the description of line 97. x here can be either i or t in your paper.

2. It seems you did not mention how the slice network was initialized and trained. Only feature extractors and classifier implementations were mentioned in section 3.1 Implementation and Metrics. I suppose you also learned the weights of the slice network. It should be specified in the paper.

Author Response

We thank the reviewer for the thoughtful comments and suggestions. We were able to address all the comments and revise the manuscript accordingly. Below are the point-by-point responses.  Please feel free to let us know if anything we can further provide or clarify. We hope with the revision, the manuscript is now suitable to be published in Electronics.

1. It is better to use the same symbol I_x for Figure 2 and the description of line 97. x here can be either i or t in your paper.

Response: We thank the reviewer for the suggestion. We have corrected the symbol accordingly (Lines 110 and 114).

2. It seems you did not mention how the slice network was initialized and trained. Only feature extractors and classifier implementations were mentioned in section 3.1 Implementation and Metrics. I suppose you also learned the weights of the slice network. It should be specified in the paper.

Response: The slice network was initialized by Kaiming initialization. The parameters of the slice network are trainable, and we learned the weights of the slice network in the model training (Lines 120-122).

Reviewer 3 Report

The paper proposes an approach for converting the 3D brain image to a 2D fused image for better performance of machine learning classifiers. The proposed methodology is evaluated on a benchmark ADNI dataset with good results. The paper is well prepared and relevant, but it needs to be improved before it could be considered for publication.

Comments:

1.       The authors should clearly state and outline the novelty of this study and its contribution to the research field (in the introduction section).

2.       Only a handful of previous studies are analyzed in detail. More recent related works should be discussed, see doi:10.1259/bjr.20211253, doi:10.3390/s22030740, doi:10.1007/978-3-030-96308-8_27.

3.       Figure 1 presents more than architecture of CNN. It looks like a full workflow. Revise the caption accordingly.

4.       Eq. 3: des the loss function has a regularization term?

5.       Why you stop training at 150 epochs? How the threat of overfitting during model training was handled?

6.       The AUC (ROC?) curves are mentioned in Line 157 but never presented. Present ROC curves.

7.       Also present and discuss the confusion matrices for comparison.

8.       Lines 249-250 claim that the proposed methodology has “significantly increased the efficiency of Alzheimer’s disease classification”. To prove this statement, perform the statistical analysis and present its results. Support by statistical testing.

Author Response

We thank the reviewer for the thoughtful comments and suggestions. We were able to address all the comments and revise the manuscript accordingly. We also added two more figures (Fig. 5 and 6) to respond to the reviewer’s comments. Below are the point-by-point responses. Please feel free to let us know if anything we can further provide or clarify. We hope with the revision, the manuscript is now suitable to be published in Electronics.

1. The authors should clearly state and outline the novelty of this study and its contribution to the research field (in the introduction section).

Response: We thank the reviewer for the comment. We added our contributions as pullet points in the Introduction (Lines 76-85).

2. Only a handful of previous studies are analyzed in detail. More recent related works should be discussed, see doi:10.1259/bjr.20211253, doi:10.3390/s22030740, doi:10.1007/978-3-030-96308-8_27.

Response: We thank the reviewer for the suggestion. We have added the references into the  Introduction (Lines 50-53).

3. Figure 1 presents more than architecture of CNN. It looks like a full workflow. Revise the caption accordingly.

Response: We have revised the caption accordingly.

3. 3: des the loss function has a regularization term?

Response: The loss function does not have a regularization term.

4. Why you stop training at 150 epochs? How the threat of overfitting during model training was handled?

Response: We set the epoch equal 150 to ensure all the models' convergence. We use cross-validation and weight decay to handle the overfitting during the model training.

5. The AUC (ROC?) curves are mentioned in Line 157 but never presented. Present ROC curves.

Response: We added a new figure (Figure 5) to present ROC curves. 

6. Also present and discuss the confusion matrices for comparison.

Response: We added a new figure (Figure 6) to present the confusion matrix.

7. Lines 249-250 claim that the proposed methodology has “significantly increased the efficiency of Alzheimer’s disease classification”. To prove this statement, perform the statistical analysis and present its results. Support by statistical testing.

Response: We thank the reviewer for the suggestion. We performed the statistical analysis on the epoch time between LWP and 3D CNN models and present the results (Lines 207-208).

Round 2

Reviewer 3 Report

The manuscript has been revised well. The quality has been improved. I recommend to accept.