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.
The paper is now suitable for publication. Congratulations.
[# PeerJ Staff Note - this decision was reviewed and approved by Shawn Gomez, a PeerJ Section Editor covering this Section #]
Dear Authors,
Thank you for the revised submission of your manuscript "Deep feature fusion for melanoma detection using ResNet-VGG networks with attention mechanism". Reviewer 1 acknowledges substantial improvements in clarity, methodological explanation, and experimental rigor. The hybrid architecture is now better articulated, and the inclusion of t-SNE visualizations and an ablation study reinforces the validity of your approach.
As a final enhancement, please consider adding a brief discussion on the computational complexity or inference time of your model compared to baselines.
Once this minor point is addressed, your manuscript will be ready for acceptance.
The revised manuscript demonstrates significant improvement in clarity, structure, and language. The authors have adequately addressed prior concerns related to grammatical inconsistencies and have rephrased multiple sections for improved readability, as visible in the tracked changes. They have also expanded the background and cited more relevant literature to establish context.
The experimental design now appears more robust. The authors have provided a clearer explanation of the hybrid architecture combining ResNet50 and VGG16 with channel and spatial attention mechanisms. They also clarified the data augmentation methods and hyperparameter settings, which were previously ambiguous.
The results are clearly presented and show significant improvement over baseline and state-of-the-art methods. Evaluation metrics such as accuracy, AUC, sensitivity, and specificity are consistently reported. The authors also provide a thoughtful t-SNE visualization to support feature discriminability, and the ablation study strengthens the validity of their attention mechanism.
(a) I appreciate the authors’ effort in thoroughly addressing reviewer comments.
(b) The inclusion of a detailed ablation study adds strong technical merit.
(c) Consider adding a small paragraph in the discussion comparing the computational complexity or inference time of the proposed model with baseline models.
**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.
**Language Note:** The review process has identified that the English language must be improved. PeerJ can provide language editing services - please contact us at [email protected] for pricing (be sure to provide your manuscript number and title). Alternatively, you should make your own arrangements to improve the language quality and provide details in your response letter. – PeerJ Staff
The manuscript is generally well-structured, and the content is logically organized. The introduction provides a comprehensive background on melanoma detection and the challenges associated with current diagnostic methods. The authors reference relevant literature, including key works in CNNs, attention mechanisms, and hybrid models.
However, there are several language and editorial issues that need attention:
(a) Language and Grammar: The manuscript contains numerous grammatical errors and awkward phrases (e.g., “CA-SLNet were evaluated,” “han’gs bypass connections,” “data amplification mechanism”). A professional proofreading or editing is strongly recommended to improve clarity and readability.
(b) Figure Captions and Visuals: Figures such as Grad-CAM visualizations and ROC curves are referenced but not clearly presented in this format. Ensure that all figures are well-captioned, high-resolution, and directly support the text.
(c) Reference Formatting: Several references (e.g., Athina 2022 from Kaggle) need proper academic formatting, and citation styles should conform to journal standards.
The experimental design is sound, and the authors have clearly described the use of ResNet50 and VGG16 architectures in combination with channel and spatial attention modules.
Please improve on the following section:
(a) No assessment of model inference time or hardware efficiency important for clinical deployment.
(b) Mention the statistical significance of the testing and show some.
The results are clearly presented and strongly support the authors’ claims. The proposed model achieves higher accuracy than competing methods across all datasets, and ablation studies further validate the contribution of attention mechanisms and dual backbone design.
However, a few concerns remain:
(a) False negatives are mentioned as an issue on ISIC-2017, but no detailed analysis or visualization of misclassified cases is provided.
(b) The interpretability (e.g., Grad-CAM heatmaps) is discussed but not shown or explained deeply enough. Clinical applications benefit from understanding why the model makes its decisions.
Suggested Improvements:
(a) Professional proofreading for grammar and clarity.
(b) Highlight limitations more transparently, particularly:
(i) Higher false negative rate on ISIC-2017.
(ii) Computational complexity of dual-backbone networks.
(c) Include interpretability visuals and analysis (Grad-CAM, saliency maps).
(d) Quantify dataset balance, misclassification patterns, and discuss mitigation.
This is a promising and well-executed study with high potential, but it needs improvements in writing quality, figure clarity, and interpretability analysis to meet PeerJ’s publication standards.
The manuscript lacks sufficient detail in key areas. The explanation of how the ResNet50 and VGG16 architectures are integrated with the attention mechanisms is too vague and should be expanded with a clear technical explanation of how channel and spatial attention contribute to feature selection and spatial focus. The dataset descriptions are minimal, and critical information such as the number of images, class distribution, and preprocessing steps should be included to enhance the transparency and reproducibility of the study. The evaluation metrics are limited to accuracy and AUC, and additional metrics such as precision, recall, F1-score, and confusion matrices should be provided to offer a more comprehensive assessment of model performance. The comparison with state-of-the-art methods is not sufficiently detailed. A more thorough comparison using specific metrics and performance analysis is required to highlight the strengths and weaknesses of the proposed method. The false negatives observed in the ISIC-2017 dataset need further investigation, and solutions or strategies to mitigate this issue should be discussed. Although clinical relevance is mentioned, the manuscript should provide more detailed insights into the practical implementation of the model, including computational efficiency and integration into clinical workflows. Furthermore, the manuscript lacks information on the model's training parameters, hyperparameters, and the computing environment, all of which are necessary to ensure reproducibility. Finally, the manuscript should address the limitations of the model and discuss potential improvements, such as using transfer learning or handling dataset biases.
The manuscript lacks details on preprocessing steps, such as normalization, augmentation, and handling class imbalance.
The dataset splitting process (training, validation, test sets) is not described.
Only accuracy and AUC are reported; additional metrics like precision, recall, F1-score, and confusion matrices are needed.
The comparison with state-of-the-art models is too vague; specific metrics should be provided.
False negatives in the ISIC-2017 dataset are not thoroughly analyzed; reasons and solutions should be addressed.
The manuscript does not provide sufficient information on training parameters (e.g., learning rate, batch size, optimizer).
The computing environment (hardware, software) is not mentioned; including this is necessary for reproducibility.
The manuscript does not provide a detailed explanation of how the attention mechanisms contribute to improving the model’s performance, which affects the validity of the findings. Clear justification for why channel and spatial attention are effective in this context is needed.
The false negative results in the ISIC-2017 dataset were not sufficiently analyzed, raising concerns about the model's robustness. A deeper exploration of these errors and potential solutions would strengthen the validity of the findings.
The manuscript relies on accuracy and AUC as the primary evaluation metrics but neglects other critical metrics such as precision, recall, and F1-score. Including these will provide a more holistic assessment of the model’s performance, especially for clinical applications where false positives and false negatives can have significant consequences.
There is insufficient discussion on the potential biases within the datasets, such as class imbalance, and how these might affect the findings. Addressing these biases would help ensure the robustness and generalizability of the model.
The model’s reproducibility is not fully ensured due to a lack of detailed information on training parameters, dataset splits, and computational environment. Providing this information will enhance the credibility and validity of the findings.
The manuscript could benefit from a more thorough discussion on the real-world impact of this work, particularly regarding its implementation in clinical diagnostic workflows and computational requirements.
It is recommended to cite recent literature on similar approaches in melanoma detection and computer vision in medical imaging to align the manuscript with current advancements in the field.
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.