Identifying Diabetic Retinopathy in the Human Eye: A Hybrid Approach Based on a Computer-Aided Diagnosis System Combined with Deep Learning

Round 1

Reviewer 1 Report

3.1. Subsection

- Is this the exact title of the subsection?

- Line 187: "It is a sizable dataset that is openly accessible and was first presented in a Kaggle competition." - provide a reference to the Kaggle data set;

section 3.4:

- A more detailed description of the structure and operation of ResNet50.4. would be useful,

Discussion:

- two subsections should be emphasized: limitations of the own study and directions for further research

Supplementary Materials: The following supporting information can be downloaded at: xxx - this should be completed

Author Response

3.1. Subsection

- Is this the exact title of the subsection?

ŞYA (corresponding author): Agree, corrected with the DR dataset

- Line 187: "It is a sizable dataset that is openly accessible and was first presented in a Kaggle competition." - provide a reference to the Kaggle data set;

ŞYA: Agree, added reference as listed [37]

section 3.4:

- A more detailed description of the structure and operation of ResNet50.4. would be useful,

ŞYA: Agree as well, regarding our discussion on ResNet architectures. In our manuscript to mention of ResNet50, it's essential to note that the numerical notations in architecture names, exemplified by "ResNet50," typically denote the network's depth or block structure.

Discussion:

- two subsections should be emphasized: limitations of the own study and directions for further research

ŞYA: This entails a thorough examination of the constraints or shortcomings within the study's scope and proposing potential avenues or topics for further investigation and development in subsequent research endeavors, given in the Conclusion part.

Supplementary Materials: The following supporting information can be downloaded at: xxx - this should be completed

ŞYA: This part will be completed after the article is uploaded to the relevant data provider during the printing phase. Please do not see it as a deficiency.

Author Response File: Author Response.docx

Reviewer 2 Report

In the article, a detailed study on the hybrid deep learning approach for the assessment of different diabetic retinopathy lesions in human eyes was proposed. This study addressed the challenges of imbalanced datasets, incorrect annotations, and lack of sample images, and compared state-of-the-art methods to improve the performance of deep learning models in this scenario. However, to increase the readability of the article, further clarification on certain points can be provided. Here are my specific comments:

(1) Lack of clear research background and motivation: In the introduction, a clearer elucidation of the challenges in diabetic retinopathy diagnosis and the limitations of current methods is needed to better highlight the importance and innovation of the study.

(2) Insufficient explanation of data availability: Although it was mentioned that the data can be provided upon request, specific explanations about data availability, such as data sources, formats, and scale, were not provided, which could impact the reproducibility and validation for other researchers.

(3) Inadequate explanation and discussion of results: In the results and discussion section, a deeper explanation of the model's performance advantages, limitations, comparisons with existing studies, and a more detailed explanation of SHAP analysis results are needed.

(4) Detailed description of image processing methods: A more detailed description of the image processing methods, particularly the challenges in identifying vessels and lesions, as well as the specific methods and innovations adopted by the authors, could be provided.

Author Response

In the article, a detailed study on the hybrid deep learning approach for the assessment of different diabetic retinopathy lesions in human eyes was proposed. This study addressed the challenges of imbalanced datasets, incorrect annotations, and lack of sample images, and compared state-of-the-art methods to improve the performance of deep learning models in this scenario. However, to increase the readability of the article, further clarification on certain points can be provided. Here are my specific comments:

(1) Lack of clear research background and motivation: In the introduction, a clearer elucidation of the challenges in diabetic retinopathy diagnosis and the limitations of current methods is needed to better highlight the importance and innovation of the study.

ŞYA (corresponding author): Agreed, added Lines 41-60 to mention challenges in DR diagnosis and why improvement is needed. Added Lines 77-90 as well to point to limitations imposed by datasets that might hinder the system's capacity to encounter and adapt to the wide array of real-world clinical variations.  

(2) Insufficient explanation of data availability: Although it was mentioned that the data can be provided upon request, specific explanations about data availability, such as data sources, formats, and scale, were not provided, which could impact the reproducibility and validation for other researchers.

ŞYA: R2 mentioned that data should be shared and other researchers should benefit from this development, and he is right in this regard. However, the data is open source and referenced. Another thing R2 is right about is that scripts should also be shared, and a correction has been made about this issue in line 526. When the manuscript is published, the scripts will be shared from an open data-sharing application, as noted by another referee. The reason why "xxx" is written in this section is that the editor has not yet given any guidance regarding which data provider the journal uses. After it is done, that part will be corrected immediately.

(3) Inadequate explanation and discussion of results: In the results and discussion section, a deeper explanation of the model's performance advantages, limitations, comparisons with existing studies, and a more detailed explanation of SHAP analysis results are needed.

ŞYA: Although Reviewer2 (R2) is right in this regard, detailed interpretation of the findings regarding the SHAP application is given between lines 359-381. However, the advantages regarding the performance of the model are adequately given at different points throughout the text. Additionally, information about the limit values of the functional application is given in lines 450-459 as agreed with R2. This information was added to lines 460-470 when the manuscript was revised with R2's comments.

(4) Detailed description of image processing methods: A more detailed description of the image processing methods, particularly the challenges in identifying vessels and lesions, as well as the specific methods and innovations adopted by the authors, could be provided.

ŞYA: Throughout the text, the advantages of CNN-based image processing over other processing are given under different headings. Accordingly, presenting the SHAP application in a hybrid way has enabled detailed information flow for image processing. As a specific method group, Reviewer has requested innovations from us, but currently the entire set of processes has an innovation in a completely hybrid form.

However, the definition of vessels and lesions symbolizes a different data filter, data focus solution, and a different solution in which the destruction caused by DR will be described as noise. For this reason, it is shared between lines 450-459 that a different methodology should be followed regarding veins.

Author Response File: Author Response.pdf

Reviewer 3 Report

The paper addresses the challenge of diagnosing diabetic retinopathy, emphasizing the role of deep learning in utilizing medical imagery to identify blood vessel damage. The study evaluates the impact of issues like unbalanced datasets, incorrect annotations, a lack of sample images, and improper performance evaluation measures on the effectiveness of deep learning models. Using three benchmark datasets, the authors conduct a comprehensive comparison of various state-of-the-art approaches. The results show moderate to high precision scores for different retinopathy phases. The paper concludes by analyzing hybrid modeling, specifically combining CNN analysis and SHAP model derivation, highlighting the effectiveness of hybrid modeling strategies for automated diabetic retinopathy detection using deep learning classification models.

Author Response

ŞYA (corresponding author): Thank you for reviewing and commenting on my manuscript file.

Reviewer 4 Report

In the manuscript ' A hybrid deep learning approach based computer-aided diagnosis system for identifying different diabetic retinopathy evaluations in human-eye, ' the authors Atci et al reported their research on developing a deep learning method to diagnose diabetic retinopathy using computers. According to the authors, the method has a high detection rate and is promising as a user-friendly technique in diabetic retinopathy diagnosis. After reviewing this manuscript, I think it can be accepted as the publication of Applied Sciences after the following questions are answered.

1. In the manuscript, the authors applied normal retrinopatric images to approve the feasibility of their method. Current research has approved several more-advanced techniques to image retrinopatrics such as OCT. Can this method applied for OCT diagnosis? Could the authors give the requirement for the imaging techniques during the method application;

2. In retrinopatric imaging, a key problem come from the motion of the human-eyes, will the motion affect the application of this method? Or, does this method has the limitation by the image quality?

Author Response

In the manuscript ' A hybrid deep learning approach based computer-aided diagnosis system for identifying different diabetic retinopathy evaluations in human-eye, ' the authors Atci et al reported their research on developing a deep learning method to diagnose diabetic retinopathy using computers. According to the authors, the method has a high detection rate and is promising as a user-friendly technique in diabetic retinopathy diagnosis. After reviewing this manuscript, I think it can be accepted as the publication of Applied Sciences after the following questions are answered.

1. In the manuscript, the authors applied normal retrinopatric images to approve the feasibility of their method. Current research has approved several more-advanced techniques to image retrinopatrics such as OCT. Can this method applied for OCT diagnosis? Could the authors give the requirement for the imaging techniques during the method application;

ŞYA (corresponding author): OCT can provide very good results in detecting DR lesions under current conditions. However, OCT is used to measure the progression of the DR level of diagnosed patients, in addition to being a very expensive method rather than being used in the diagnosis of potential patients. However, opportunities for OCT imaging are universally limited, taking a simple retinopathy image and processing it immediately offers a much more effective solution for detecting DR.

2. In retrinopatric imaging, a key problem come from the motion of the human-eyes, will the motion affect the application of this method? Or, does this method has the limitation by the image quality?

ŞYA: Especially headings 3.2, 3.3. and in 3.5 this issue was indirectly touched upon. Image quality represents the eyepiece being at full focus under optimum conditions, but there are very few such perfect examples in the dataset. As mentioned in the comment, in addition to the moving eye focus point, incorrect shooting, camera errors, camera version differences, undeveloped camera features, etc. Filters and CNN applications have been developed by taking into account many features, and SHAP application for location-based detection has been implemented through standardized examples.

Round 2

Reviewer 2 Report

The paper can be accepted.

Author Response

ŞYA (corresponding author) : Thanks for all your contributing comments.