All reviews of published articles are made public. This includes manuscript files, peer review comments, author rebuttals and revised materials. Note: This was optional for articles submitted before 13 February 2023.
Peer reviewers are encouraged (but not required) to provide their names to the authors when submitting their peer review. If they agree to provide their name, then their personal profile page will reflect a public acknowledgment that they performed a review (even if the article is rejected). If the article is accepted, then reviewers who provided their name will be associated with the article itself.
Thank you for your valuable contribution!
[# PeerJ Staff Note - this decision was reviewed and approved by Mehmet Cunkas, a PeerJ Section Editor covering this Section #]
Now, the improvement is noticeable in the paper.
The authors addressed all questions and reservations.
Table 7 shows now the computational gain of the proposed method. Furthermore, the modified Table 6 became more informative.
Figure 7 presents what the model is looking for within the same class.
Now, the manuscript has reached certain scientific values that deserve to be published.
Accepted
Accepted
Accepted
Accepted
Please address the criticisms and comments from Reviewer 1 thoroughly before resubmitting your manuscript.
I appreciate the authors’ efforts in responding to my comments. However, some of the major concerns raised in the first review have not been adequately addressed. Specifically:
1- The comparisons with recent state-of-the-art hybrid models presented in Table 5 of the manuscript lack essential information (accuracy, precision, number of samples used, data name, experimental protocol, etc.). This makes your comparison less meaningful.
2- The authors claim to have added the APTOS dataset, but in all tables and figures, the results and interpretation related to this data are missing (except figure 10, which mentions the results on both datasets) (Table 5 with the proposed model only, not with the other models).
3- The content and aim of Table 2 are not clear; more explanation is needed about it.
Using GRAD-CAM, Grad-CAM helps with the visual interpretation of decision models, but Figure 7 does not really show this. It is recommended to show some DR images of the same class but with different decisions to see where the proposed model is looking/lacking.
Also, the authors claim to have evaluated the computational efficiency of the proposed system, but no experiments or studies are available to address that.
The authors claim to have added the APTOS dataset, but in all tables and figures, the results and interpretation related to this data are missing (except figure 10, which mentions the results on both datasets) (Table 5 with the proposed model only, not with the other models).
-
-
-
The authors have addressed all the comments.
**PeerJ Staff Note:** It is PeerJ policy that additional references suggested during the peer-review process should only be included if the authors agree that they are relevant and useful.
**PeerJ Staff Note:** Please ensure that all review and editorial comments are addressed in a response letter and that any edits or clarifications mentioned in the letter are also inserted into the revised manuscript where appropriate.
This paper presents a well-structured system called RetinoNet for automated diabetic retinopathy (DR) classification.
The authors integrated a long chain of pre-processing steps, EfficientNet-B0, FPN, GAP and statistical validation.
The obtained results (96.8% accuracy, 97.5% AUC) are good, and the methodology addresses key challenges like small datasets and bias.
However, clarity issues, missing implementation details, and dataset limitations need addressing.
RetinoNet combines EfficientNet-B0 for hierarchical feature extraction with a Feature Pyramid Network (FPN) to capture multi-scale lesions (e.g., microaneurysms, hemorrhages). This integration improves detection of both small and large lesions, addressing a key challenge in DR diagnosis
The model uses the Messidor dataset (1,200 images across 4 severity classes), augmented via rotation, flipping, and shearing. Code and curated data are publicly available on GitHub and Kaggle.
The effective fusion of EfficientNet-B0 (feature extraction), FPN (multi-scale lesion detection), and GAP (overfitting reduction) is well-justified.
The data augmentation (flipping/rotation) is basic. So, it is recommanded to use advanced augmentation (e.g., generative adversarial networks) or test on more than one external dataset to strengthen claims of robustness.
The authors sould resond to the following points:
1- Training on 1,200 images (post-augmentation) risks overfitting despite augmentation, so an external validation on larger datasets (e.g., EyePACS or APTOS) is required.
2- While outperforming older models, direct comparisons to state-of-the-art hybrids (e.g., FPN + ResNet34 or EFPN) are absent.
3- Limited analysis of false positives in mild/moderate cases, which are critical for early diagnosis.
4- The class imbalance problem should be discussed.
5- Cross-dataset validation to assess generalizability is also important due to the small data used.
6- This work needs broader validation and transparency in methodology to solidify its clinical relevance.
7- The proposed framework requires further optimization for real-world scalability.
8- How are foreground/background seeds determined? eq. 2
9- ANOVA results (Table 3) lack critical values (F-statistic, p-values). Claims of "no bias" need fuller statistical evidence.
N/A
N/A
N/A
1. The choice of EfficientNet-B0 is not justified; compare its performance with deeper variants or transformer-based backbones to support its selection.
2. The impact of FPN and GAP is not isolated; include an ablation study to quantify their individual contributions to accuracy and generalization.
3. The dataset split strategy is unclear; apply k-fold cross-validation or stratified sampling to prevent data leakage and validate robustness.
4. With only 461 labeled images, the model is at high risk of overfitting; evaluate on an external dataset (e.g., EyePACS) to support generalizability.
5. No statistical tests are applied when comparing with other models; include significance testing (e.g., p-values, confidence intervals) to validate claims.
6. Prefetching and dropout-based optimizations are qualitatively mentioned; provide runtime benchmarks and resource usage to quantify computational gain.
7. ANOVA applied to image quality metrics is insufficient; correlate enhancement metrics with classification outcomes to confirm their relevance.
8. The model lacks explainability; include Grad-CAM or equivalent visualizations to show lesion-level focus and support clinical trust.
9. The effect of data augmentation is not evaluated; perform a sensitivity analysis to measure how each technique impacts classification performance.
10. Cite more recent deep learning approaches (2023–2025), especially works on transformer architectures and contrastive/self-supervised learning for DR.
1. Zeng, Y., Zhang, R., Zhang, H., Qiao, S., Huang, F., Tian, Q.,... Peng, Y. (2025). GCCNet: A Novel Network Leveraging Gated Cross-Correlation for Multi-View Classification. IEEE Transactions on Multimedia, 27, 1086-1099. doi: 10.1109/TMM.2024.3521733
2. Zhang, R., Lin, Y., Wu, Y., Deng, L., Zhang, H., Liao, M.,... Peng, Y. (2024). MvMRL: a multi-view molecular representation learning method for molecular property prediction. Briefings in Bioinformatics, 25(4), bbae298. doi: 10.1093/bib/bbae298
3. Zhang, Z., Liu, Z., Martin, A., & Zhou, K. (2023). BSC: Belief Shift Clustering. IEEE Transactions on Systems, Man, and Cybernetics: Systems, 53(3), 1748-1760. doi: 10.1109/TSMC.2022.3205365
4. Huang, B., An, H., Chu, J., Ke, S., Ke, J., Qiu, Y.,... Liu, W. (2025). Glucose-Responsive and Analgesic Gel for Diabetic Subcutaneous Abscess Treatment by Simultaneously Boosting Photodynamic Therapy and Relieving Hypoxia. Advanced Science, e02830. doi: https://doi.org/10.1002/advs.202502830
5. Wu, Z., Sun, W., & Wang, C. (2024). Clinical characteristics, treatment, and outcomes of pembrolizumab-induced uveitis. Investigational New Drugs, 42(5), 510-517. doi: 10.1007/s10637-024-01464-w
6. Liang, J., He, Y., Huang, C., Ji, F., Zhou, X.,... Yin, Y. (2024). The Regulation of Selenoproteins in Diabetes: A New Way to Treat Diabetes. Current Pharmaceutical Design, 30(20), 1541-1547. doi: https://doi.org/10.2174/0113816128302667240422110226
7. Song, W., Wang, X., Guo, Y., Li, S., Xia, B.,... Hao, A. (2024). CenterFormer: A Novel Cluster Center Enhanced Transformer for Unconstrained Dental Plaque Segmentation. IEEE Transactions on Multimedia, 26, 10965-10978. doi: 10.1109/TMM.2024.3428349
8. Yu, X., Luan, S., Lei, S., Huang, J., Liu, Z., Xue, X.,... Zhu, B. (2023). Deep learning for fast denoising filtering in ultrasound localization microscopy. Physics in Medicine & Biology, 68(20), 205002. doi: 10.1088/1361-6560/acf98f
9. Sun, J., Yu, X., Li, W., Jia, B., Shi, D., Song, Y.,... Jiang, C. (2025). Real-time accurate detection and analysis of breath acetone using CRDS: Toward metabolic dynamic monitoring and potential application. Sensors and Actuators B: Chemical, 433, 137422. doi: https://doi.org/10.1016/j.snb.2025.137422
10. Wei, Y., Xu, S., Wu, Z., Zhang, M., Bao, M.,... He, B. (2024). Exploring the causal relationships between type 2 diabetes and neurological disorders using a Mendelian randomization strategy. Medicine, 103(46), p e40412. doi: 10.1097/MD.0000000000040412
11. Wang, G., Ma, Q., Li, Y., Mao, K., Xu, L.,... Zhao, Y. (2025). A skin lesion segmentation network with edge and body fusion. Applied Soft Computing, 170, 112683. doi: https://doi.org/10.1016/j.asoc.2024.112683
12. Hu, T., Meng, S., Liu, C., Fang, W., Xia, Z., Hu, Y.,... Xia, X. (2025). LCN2 deficiency mitigates the neuroinflammatory damage following acute glaucoma. Theranostics, 15(7), 2967-2990. doi: 10.7150/thno.104752
13. Wang, S., Dong, B., Xiong, J., Liu, L., Shan, M., Koch, A. W.,... Gao, S. (2025). Phase manipulating Fresnel lenses for wide-field quantitative phase imaging. Optics Letters, 50(8), 2683-2686. doi: 10.1364/OL.555558
14. NIMEQ-SACNet: A novel self-attention precision medicine model for vision-threatening diabetic retinopathy using image data
15. EdgeSVDNet: 5G-enabled detection and classification of vision-threatening diabetic retinopathy in retinal fundus images
**PeerJ Staff Note:** It is PeerJ policy that additional references suggested during the peer-review process should only be included if the authors are in agreement that they are relevant and useful.
All text and materials provided via this peer-review history page are made available under a Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.