A Wrapped Approach Using Unlabeled Data for Diabetic Retinopathy Diagnosis
, Sang-Yong Rhee 5 and Yong Soo Kim 6,*Round 1
Reviewer 1 Report
This paper proposes a proposed self‐supervised Learning approach to overcome a shortage of sufficient data in diabetic retinopathy detection. My comments are as follows:
1- In the abstract section, I would suggest that the author should provide the point and quantitative advantages of detection results.
2- The main contributions of this paper should be further summarized and clearly demonstrated.
3- The limitations of this work should be discussed in the experimental result and discussion section.
4- I recommend that the authors should present some visualization results using Gradient-weighted Class Activation Mapping (Grad-CAM) to understand which parts of the image are most important for classification
5- Some new references should be added to improve the literature review—for example, https://doi.org/10.3390/app12115500; https://doi.org/10.3390/diagnostics12081975;https://doi.org/10.1007/s00247-022-05510-8.
Overall this can be a useful work, and some interesting aspects are presented. I recommend that the authors should revise their paper according to the above-addressed points.
Author Response
We have revised as an attached file.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
The manuscript proposes the implementation of a deep learning-based binary classification task for the presence and absence of diabetic retinopathy using a labeled and imbalanced dataset. The approach which overcomes this biased DR detection model is to add an auxiliary procedure to the target task that identifies diabetic retinopathy (DR) using supervised learning. The added process uses unlabeled data to pretrain the model that first learns features from data using self-supervised or semi-supervised learning, and then the pretrained model is transferred with the learned parameter to the target model.
1) At the end of the abstract, it will be more intuitive and convincing to illustrate the qualitative results of a large number of experiments for verifying the superiority and effectiveness.
2) The authors ignore some relevant papers. For example, fusion methods that have been published in 2022 please see ("Brain tumor segmentation based on the fusion of deep semantics and edge information in multimodal MRI" , Information Fusion ,Volume 91, Pages 376-387.) The authors should compare their method with it carefully.
3) At the first paragraph of Section2.2.1, the thesis mentions that“After pretraining, the learned parameters were then used to initiate the encoder in the module ‘2’, which was the target model of performing DR detection trained or fine‐tuned on the labeled data using supervised learning”. How is the target model specifically designed? What are the optimizations?
4) The explanation of equation 1-2 is not enough, What the symbol on the right side of the equation represents puzzles the readers.
5) Figure 3 that represents examples of unusable fundus image data could be enlarged to express the results more clearly and intuitively, without the appendix.
6) Make sure your conclusions appropriately reflect on the strengths and weaknesses of your work, how others in the field can benefit from it, and thoroughly discuss future work.
Author Response
We have revised as an attached file.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
The Authors have addressed all of my concerns with the original manuscript. The revised manuscript is ready for publication.
Reviewer 2 Report
All my concerned problems are revised.
It is suggested to be accepted.
