Enhancing Cataract Detection through Hybrid CNN Approach and Image Quadration: A Solution for Precise Diagnosis and Improved Patient Care

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

In this study, the term segmentation was misused. What the authors suggest means is quadration of an image. In deep learning, segmentation means different tasks, so caution is required.

Although the authors' approach is not very new, performance improvement can be expected by giving the effect of their own ensemble model, and it is meaningful that ui was well made. These advantages should be more emphasized.

Techniques for diagnosing ophthalmic diseases with existing fundus should be further reviewed. Recently, a technology that makes insufficient data into gan has also appeared. (e.g., Automated detection of crystalline retinopathy via fundus photography using multinational networks)

 

These various techniques should be reviewed more comprehensively.

Cataracts are affected by various factors as well as age, and diagnostic criteria may be unclear. Describe the shortcomings of the study that did not review LOCS iii.

 

Comparative analysis should be performed.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

This manuscript proposed a hybrid CNN approach for cataract diagnosis. DenseNet121 was applied as the base framework and five different models were trained based on the whole image and 4 subdivided images. Evaluation was done on two data sets and comparison with other existing deep learning based methods.

  

Major comments:

1. The manuscript is too long. Some techniques that have already been widely used can be introduced more briefly.

2. The contribution of the manuscript is limited. And there’s no evaluation of the proposed method’s operating efficiency, is it better than existing methods?

3. Chapter 3.2.4 does not explain clearly why the random rotation is 30°, not a random value within a specific range. And why ‘zoom in’ was not done in different scales.

4. It’s not described clearly whether the training, validation and testing set of the 5 CNN networks exactly the same.

5. It’s not described whether data augmentation was also done on the subdivided images.

6. The second column of Figure 11 is misleading. Is it cropped from a ROI of the original image?

 

Minor comments:

1. Inconsistent fonts used in tables. And some format issues exist.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The authors have clearly revised the paper. There are several advantages to this method.

In Figure 6, make clear whether the augmentation processes were applied to each model.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

This manuscript proposed a hybrid CNN approach for cataract diagnosis. DenseNet121 was applied as the base framework and five different models were trained based on the whole image and 4 subdivided images. Evaluation was done on two data sets and comparison with other existing deep learning based methods.

 

Minor comments:

1. Words in Table 9 still have different fonts.

2. What’s the input size of Model Top 1, Model Top 2, Model Bottom 1 and Model Bottom 2? Is it 224/2 × 224/2, or resize to the same input size as Model 1?

Author Response

Please see the attachment

Author Response File: Author Response.pdf