Deep Feature Vectors Concatenation for Eye Disease Detection Using Fundus Image

Round 1

Reviewer 1 Report

The article was well written, bringing a diagnose technique for diabetic retinopathy (DR), a disease with an increasing number of cases. This work can increase the efficiency of ophthalmologists to diagnose the DR and treat it, being of great relevance since this illness can result in vision loss.

The authors could describe the pooling steps of Figure 7's flowchart.

How the authors concluded the paragraph below:
We use MLP with 1024 hidden layers. Thus, the model built can see better from two points of view of feature learning with the DenseNet121 and InceptionResNetV2 models to recognize the pattern of lesions in the fundus images.

The GPU board specifications need more details like the number of cores and teraflops performance, and so on.

The authors should explain the cropping images process (if they cropped images manually or using an automated process).

 

Author Response

Thank you for the comments and suggestions. Here I attach responses from reviewers' comments based on discussions with other authors, grammatical consultations and further literature study. I hope there is a response from the reviewer regarding the revision regarding whether or not the revision was accepted because it relates to my report responsibilities in the grant report.

Author Response File: Author Response.pdf

Reviewer 2 Report

The presented paper entitled "Deep Feature Vectors Concatenation for Eye Disease Detection using Fundus Image" describes a very urgent problem concerning diabetes and its early diagnosis. The applied methods of fundus image classification significantly support the work of medical specialists, accelerating the diagnosis.

However, the article itself has several minor flaws. Firstly, the motivation for using such classifiers is not sufficiently explained. Secondly, the depiction of the architecture of the classifiers by means of pictures is not entirely clear. Thirdly, the comparison with other previously known classification methods should, in my opinion, be done on the same set of data. The last thing is a strong need to rewrite the introduction. There is a lot of repetition and it deviates a lot from the conciseness of the article. To sum up, the article presents an interesting topic and its solution is valid, however it requires the authors to improve a few important issues mentioned before. 

Author Response

Thank you for the comments and suggestions. Here I attach responses from reviewers' comments based on discussions with other authors, grammatical consultations and further literature study. I hope there is a response from the reviewer regarding the revision regarding whether or not the revision was accepted because it relates to my report responsibilities in the grant report.

Author Response File: Author Response.pdf

Reviewer 3 Report

Fundus image is an image that captures the back of the eye (retina) which plays an important role in the detection of a disease, including diabetic retinopathy (DR). It is the most common complication in diabetics that remains an important cause of visual impairment, especially in the young and economically active age group. In patients with DR, early diagnosis can effectively help prevent the risk of vision loss. DR screening was performed by an ophthalmologist by analysing the lesions on the fundus image. However, the increasing prevalence of DR is not proportional to the availability of ophthalmologists who can read fundus images. It can lead to delayed prevention and management of DR. Therefore, there is a need for an automated diagnostic system as it can help ophthalmologists increase the efficiency of the diagnostic process. This paper provides a deep learning approach with the concatenate model for fundus image classification with three classes: no DR, non-proliferative diabetic retinopathy (NPDR) and proliferative diabetic retinopathy (PDR). This paper is well written and organized. Some suggestions are as follows:

 

[1] In the Abstract, “DR screening was performed by an ophthalmologist by analysing the lesions on the fundus image.”, should be “DR screening was performed by an ophthalmologist by analyzing …”

”

[2] The title of Fig. 3, “Proposed method design” should be “The design of the proposed method”. In figure 1. “A fundus image” should be “Fundus images”. Too many syntax errors appeared. The author is requested to carefully modify the language of the paper to improve its language quality.

 

[3] If some robust classifiers are used in the last layer instead of the softmax, the accuracy may be further improved. This point can be discussed in the future work, and some robust classifiers can be referred as follows:

 

Transfer Learning with Label Noise[J]. 2017

Granular Ball Sampling for Noisy Label Classification or Imbalanced Classification[J]. IEEE Transactions on Neural Networks and Learning Systems, 2021.

mCRF and mRD: Two Classification Methods based on a Novel Multiclass Label Noise Filtering Learning Framework [J]. IEEE Transactions on Neural Networks and Learning Systems. 2021

Author Response

Thank you for the comments and suggestions. Here I attach responses from reviewers' comments based on discussions with other authors, grammatical consultations and further literature study. I hope there is a response from the reviewer regarding the revision regarding whether or not the revision was accepted because it relates to my report responsibilities in the grant report.

Author Response File: Author Response.pdf

Reviewer 4 Report

The manuscript presents the implementation of well researched machine learning algorithms to detect and classify retina disease in diabetic or pre-diabetic patients using fundus images from the eye. The authors describe the framework used which consists of a feature extraction algorithm (DenseNet121 or Inception-ResNet-V2) and a classification algorithm (Multi Layer Perceptron). The authors make use of a recently released data set, namely the Dataset for Diabetic Retinopathy (DDR), which was released to the public in 2019. Both the data set and algorithms are known to the scientific community, hence the original contribution of the authors is the use of these algorithms together with the DDR data set. The authors do not explicitly and clearly state what is their main contribution and its novel aspects. The manuscript restricts itself to merely the application of the algorithms to the new data set, which has limited novel contribution to the scientific community. The presentation of the algorithm and explanation of the data set used, as well as the post-processing needed for the whole classification framework, are well explained and presented. I recommend to accept the manuscript with minor modifications. Please address the following comments:

1. As mentioned in the summary, the novel contribution from the authors presented in the manuscript is not clearly or explicitly stated. The authors need to provide a strong argument why the mere implementation of these well-known algorithms to the public data set is worth publishing. Without such statement, the paper does not provide a strong contribution to the field.
2. On line 125, the authors mention that the DDR fundus images are seperated into three channels (red, green and blue) so as to increase the contrast of the feature of the image. Each DDR fundus image has an uncertainty associated with them, since depending on the technician, the images might not capture important features from the eye. Do the authors consider the impact of uncertainty in the image to the feature extraction algorithms? What is the expected impact of uncertainty due to errors in the images?
3. In the conclusions, the authors mentioned that the data collected only contained three classes. Is the data referred in here the DDR fundus images? Such data has six categories: normal, mild NPDR, moderate NPDR, severe NPDR, PDR, and ungradable. It is the authors themselves that re-classify the data into three classes: normal, NPDR (mild, moderate, and severe), and PDR. Why do the authors not make use of a wider number of classifications?

Author Response

We thank the editor and reviewers for thoroughly reading our manuscript. We appreciate the comments that have been given in order to provide better feedback on this manuscript by reviewers.

Author Response File: Author Response.pdf

Round 2

Reviewer 3 Report

The authors have revised the paper according to the reviewers' comments.

Author Response

We thank the editor and reviewers for thoroughly reading our manuscript. We hope that with some changes to the manuscript, it will improve the quality of the manuscript that has been done.

Author Response File: Author Response.pdf