Enhancing Melanoma Diagnosis with Advanced Deep Learning Models Focusing on Vision Transformer, Swin Transformer, and ConvNeXt

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The paper examines an important problem regarding the detection of melanoma from dermoscopic images using deep learning models. The motivation is clearly outlined in the introduction with relevant statistics on rising skin cancer cases globally. However, I think the background could be strengthened by elaborating more on the limitations of existing techniques for this task. The authors could cite a few more key papers to better set the context.

1. The idea of using advanced models like ConvNeXt and Vision Transformers is logically presented, but I'd recommend providing some more intuition on why their architectural innovations are well-suited for this medical imaging application. For instance, the ability of ViT models to understand global context and long-range dependencies can be highly beneficial for distinguishing ambiguous lesions. 

2. The study design seems appropriate with a standard train-test split methodology. My main concern is with the details provided in the methods section. Model training details like batch size, number of epochs, optimizer parameters are missing which makes reproducibility impossible. I'd suggest expanding the methodology description substantially to give readers enough clarity to replicate or build on this work. 

3. The results are presented clearly in terms of various classification metrics. However, the class imbalance issue has not been addressed which could provide misleading accuracy metrics if the dataset is unbalanced. Reporting sensitivity, specificity, ROC curves would give better clinical context beyond just accuracy. 

4. The discussion/conclusion highlight the potential of these advanced models for improving melanoma diagnosis. However, translating these methods to clinical practice involves many challenges like acquisition condition, image resolution, explainability/uncertainty etc. which are not discussed. I'd recommend expanding the discussion on practical limitations and also suggesting future research directions beyond just using more data. For example, color constancy algorithm have proven to increase classification performance while standardising the image appearance (doi: 10.1016/j.cmpb.2022.107040, doi: 10.1109/JBHI.2014.2336473, doi: 10.1016/j.eswa.2023.123105). Explainability is also widely used in dermatology and could be a solid future direction in this research field (doi: 10.1016/j.media.2022.102647, doi: 10.1038/s41467-023-43095-4, doi: 10.1016/j.ejca.2022.02.025)

Comments on the Quality of English Language

-

Author Response

- Thank you for highlighting the importance of strengthening the background section. We enhanced this section by detailing the limitations of existing techniques for melanoma detection. Additional key papers were cited to better set the context and provide a comprehensive overview of the current state of the field.

- We appreciate your suggestion to provide more assessment on the suitability of ConvNeXt and Vision Transformers for medical imaging applications. We included a detailed explanation of how the architectural innovations of these models, such as the ability of ViT models to understand global context and long-range dependencies, are beneficial for distinguishing ambiguous lesions.

- Your concern about the missing training details is well noted. We expanded the methodology section to include comprehensive information on batch size, number of epochs, optimizer parameters, and other relevant training details. This ensure that our study is fully reproducible and provides clarity for other researchers to build upon our work.

- We acknowledge the need to address the class imbalance issue. We highlighted our analysis of the class distribution in our dataset and reported additional metrics such as sensitivity, specificity, and ROC curves. This provides a more complete clinical context and lessens any potential bias in the accuracy metrics.

- We agree that discussing the practical limitations and challenges of translating these methods to clinical practice is crucial. We expanded the discussion section to address issues such as acquisition conditions, image resolution, explainability, and uncertainty. Additionally, we suggested future research directions, including the use of color constancy algorithms and explainability techniques in dermatology, supported by the references you provided.

Reviewer 2 Report

Comments and Suggestions for Authors

In this article, the authors consider potential of hybrid deep learning architectures in medical image analysis, particularly for early melanoma detection. The study utilized Con- 368 vNeXt, Vision Transformer (ViT) Base-16, and Swin Transformer V2 Small (Swin V2 S),  demonstrating the efficacy of state-of-the-art techniques in enhancing diagnostic accuracy.

Еhere are the following comments:

1) Table 1. Sigmoid: Outputs between 0 and [the sentence is not complete]

2) Clauses 4.3, 4.4 and 4.5 sound identical (Proposed Deep Learning Model)

3) Figure 1 needs to be detailed and clarified. The lack of connections is inexplicable. The picture is difficult to understand. There are missing details.

4) Figure 2. Proposed methodology . You need to highlight how your methodology differs from the generally accepted one.

5) Please indicate in more detail the element of scientific novelty and more clearly highlight the features of the proposed approach in point 4.4.

6) It is necessary to specify the method of obtaining images in the dataset and compliance with all norms and rules for processing personal medical data.

In general, the work is of scientific importance, but needs correction according to comments

Author Response

- Thank you for pointing out the incomplete sentence in Table 1. We corrected the sentence to ensure it clearly conveys the intended information.

- We appreciate your observation regarding the redundancy in Clauses 4.3, 4.4, and 4.5. We revised these sections to eliminate repetition and ensure each clause distinctly contributes to the description of our proposed deep learning model.

- Your feedback on Figure 1 is valuable. We revised the figure to include detailed connections and ensure it is easy to understand. Additional annotations and explanations were provided to clarify the workflow and methodology depicted.

- We revised Figure 2 to highlight how our proposed methodology differs from generally accepted ones. This includes clear annotations and a detailed explanation of the unique aspects of our approach.

- We acknowledge the need to clearly specify the scientific novelty and unique features of our approach. In point 4.4, we explicitly outlined the innovative aspects and contributions of our study to the field of melanoma detection.

- We provided detailed information on the method of obtaining images for our dataset and ensure compliance with all norms and rules for processing personal medical data. This includes a description of ethical considerations, consent procedures, and data anonymization methods.

 

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

the authors addressed all my comments 

Reviewer 2 Report

Comments and Suggestions for Authors

The authors did a good job, took into account previous comments and improved the quality of the article.

Therefore, the article has the right to be published.