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 authors have addressed all of the reviewers' comments.
[# PeerJ Staff Note - this decision was reviewed and approved by Mehmet Cunkas, a PeerJ Section Editor covering this Section #]
Please, pay attention to the comments of Reviewer 1.
**PeerJ Staff Note:** Please ensure that all review, editorial, and staff 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.
I use it to improve the readability of my text.
This study show a major challenge found in the limited knowledge of species’ biological traits within transmission cycles, which is crucial for accurate zoonotic disease risk assessment and forecasting.
1) From an engineering perspective, implementation is feasible, though the development of entirely new algorithms may not be the main point of interest. Clinically, identifying the ecological roles of species—whether as hosts or reservoirs—enables the application of community and landscape ecology frameworks that account for species interactions and spatial dynamics, thereby improving prediction accuracy.
2) However, it remains uncertain whether genus-level identification alone provides sufficient information for planning effective control measures.
3) The practical outcome of this work is a web-based application, which holds promise for field implementation, though broader validation is needed to ensure generalizability.
-
The authors did a great job making this paper much clearer and more professional. They fixed all the scientific terms, like correctly using "acarologist" for a tick expert and properly formatting tick names. They also removed exaggerated or unclear language to make the writing more straightforward. The paper is now easier to follow because they clearly explain why they focused on the two tick types found in their region of Cyprus.
The way the study was designed is now much better and easier to understand.
The authors resolved the ambiguity of the "unidentified" class by creating a logical two-stage process: first, a binary classification between ticks and non-ticks (Experiment A), and second, a classification between the two tick genera (Experiment B).
Reproducibility is greatly enhanced by the inclusion of specific methodological details, such as the brand and model of the microscope used for image capture and a clear description of the data augmentation techniques applied during training.
The process of generating thousands of images from 35 tick specimens by capturing multiple anatomical views has also been well-clarified.
The study's results are now more believable because the authors used stronger methods to check their models' performance. They added important measures that were missing before, like the AUC score, which shows how well the models can classify the ticks. They also used a statistical test (ANOVA) to prove that the performance differences between their models are real and not just due to random chance. Because of these improvements, their conclusion that the models work well is strongly supported by the data.
The authors should be commended for their exceptional diligence in addressing all reviewer comments. The manuscript is greatly improved by the addition of new sections that thoughtfully consider the real-world context of this work. The new "Benefits and Risks of Automated Tick Identification" section adds significant value by responsibly discussing the limitations of the AI model, including the potential for false negatives and user complacency. Furthermore, the inclusion of a "Clinical or Practical Implication" section successfully highlights the tool's potential value for disease surveillance and veterinary medicine. These additions, combined with the development of the functional web application, make this a well-rounded and impactful study.
This version looks better. Thanks for fixing the review comments.
This version looks better. Thanks for fixing the review comments.
This version looks better. Thanks for fixing the review comments.
This version looks better. Thanks for fixing the review comments.
Dear Authors,
Thank you for submitting your manuscript to our journal. After careful evaluation by four reviewers, we have reached a decision.
The reviewers found the topic to be timely and relevant, particularly your application of AI for tick identification and the development of a web-based tool. However, substantial concerns were raised regarding the manuscript’s novelty, the clarity and completeness of the experimental design, the absence of key performance metrics (such as AUC and statistical significance testing), and the limited taxonomic scope.
In light of these comments, we are requesting that you revise your manuscript substantially before it can be considered for publication. Specifically, please address the following key points:
• Provide justification for the focus on only two tick genera and clarify the scope in relation to existing tools.
• Expand on methodology details (data augmentation, model selection criteria, hardware used, image acquisition).
• Improve the statistical analysis by including AUC values, validation of performance differences, and error analysis.
• Enhance clarity regarding dataset composition, the role of non-tick images, and the “unidentified” class.
• Revise ambiguous or informal language and ensure biological terminology is used accurately throughout.
• Discuss broader implications, including usability, limitations, and potential health relevance of automated tick identification.
Two reviewers requested major revisions, and two others suggested minor revisions focused on clarity and biological relevance. Taken together, we are issuing a Major Revision decision.
We invite you to submit a thoroughly revised version that carefully addresses each reviewer’s comments in a point-by-point response.
**Language Note:** PeerJ staff have identified that the English language needs to be improved. When you prepare your next revision, please either (i) have a colleague who is proficient in English and familiar with the subject matter review your manuscript, or (ii) contact a professional editing service to review your manuscript. PeerJ can provide language editing services - you can contact us at [email protected] for pricing (be sure to provide your manuscript number and title). – PeerJ Staff
This study focuses on developing a web-based application using customized CNN and pre-trained models to classify two tick genera while minimizing false predictions (e.g., misclassifying spiders as ticks) through targeted training. However, while the experimental design appears sound, the application lacks novelty, and key evaluations, such as accuracy, speed, and generalization ability, are missing. Additionally, the manuscript’s structure needs refinement. To enhance the study, the researchers could address these gaps by incorporating performance benchmarks and improving the organization of their writing.
Introduction Section:
-Line 86: Please ensure all abbreviations are fully defined upon first use.
-Lines 103–104 require supporting in-text citations.
-Lines 109–112 lack references; relevant literature should be cited here.
-Clarify Scope Discrepancy at lines 135–136, note that the TickPhone app identified 31 tick genera, yet this study focuses on only 2 genera. The rationale for this narrower scope (e.g., model simplicity, data availability, or targeted disease relevance) should be explicitly justified to avoid reader confusion.
-Microscope Specifications: The brand, model, and key specifications of the microscopes used for image capture should be clearly stated to ensure reproducibility.
-Citation Needed for Spider Data Role (Lines 178–179):
-The claim that "spider images help reduce overfitting" requires a supporting reference or empirical justification.
-Clarify the Scope of Two Genera (Hyalomma & Rhipicephalus), explain why these two genera were prioritized (e.g., disease relevance, regional prevalence, or morphological distinctiveness). In the literature Context, compare with the broader tick diversity reported by Omodior et al. (2021) to contextualize the study’s narrower focus.
-Standardized Training Views: describe the standard anatomical views (e.g., dorsal/ventral) of ticks used to train the models. This transparency would help assess how false negatives during validation were mitigated.
-Discrepancy in Sample Size (N=35 vs. 10–20 ticks per dog), clarify the total sample collection process. If N=35 refers to individual ticks, reconcile this with the stated range of 10–20 ticks per stray dog. If it refers to dogs sampled, specify the average ticks collected per dog and the total ticks analyzed
-Please give citation in lines 198-201.
-"Unidentified" Class Justification (Lines 248–252): Explain why spiders were assigned to an "unidentified" class rather than being explicitly labeled. If the model cannot confidently classify an input into any defined tick genus, how does labeling it as "unidentified" improve real-world usability?
-Practical Implications: If the model encounters a non-spider unknown sample (e.g., another arthropod), how will it be handled? Does "unidentified" strictly apply only to spiders, or is it a broader rejection class? Cite relevant literature to support this design choice.
-Training Epochs & Performance (Figure 3), epoch sufficiency, training with only 20–30 epochs may be insufficient for convergence. Consider extending epochs or implementing early stopping with a patience threshold to ensure optimal learning without overfitting. Clarify the difference between left (training) and right (validation) accuracy plots. If validation accuracy plateaus or diverges, discuss potential causes (e.g., data imbalance, insufficient augmentation).
-Model Selection Criteria for CNN, VGG16, and ResNet50, define quantitative and qualitative benchmarks for selecting the best-performing model per architecture, such as: primary Metrics for highest validation accuracy, lowest loss, or best AUC-ROC (if applicable).
-Missing AUC Analysis, critical oversight, the Area Under the Curve (AUC) is a standard metric for binary/multiclass classification, yet it’s unreported. Provide AUC values with confidence intervals to objectively compare model discrimination ability.
-Lack of novelty in AI model analysis, the discussion reiterates well-established knowledge (e.g., pre-trained models generally outperform shallow CNNs) without contributing new insights. To add value, consider: Comparing pre-trained models against lightweight architectures (e.g., MobileNet, AlexNet) to assess trade-offs between accuracy and computational efficiency. Discussing why certain models underperformed (e.g., data scarcity, architectural limitations) rather than stating expected outcomes.
-Also, an unsubstantiated proposal for an image scale experiment, the suggestion to evaluate "varying image scales" lacks empirical support, as no results from such an analysis are presented. To justify this proposal to demonstrate its impact through pilot experiments (e.g., accuracy/loss curves at different scales). And cite prior work where scale variation improved performance in similar tasks.
The paper is mostly clear and the writing is generally good. The introduction sets the stage well but i would recommend few minor suggestions.
Consistency in using "I-Tick" or "I-TickNet" , clarity on "unidentified class" like is it non-tick images or something else. It would be helpful to clearly state how the 35 ticks resulted in 19,345 images.
The design is generally sound, web app is good. However certain areas need improvements like
Explain role of 6,000 non-tick images (Lines 177-178, 225). To ignore non-ticks?
Briefly explain image generation from 35 ticks for diversity.
Results are strong, especially for VGG16. To further enhance them:
Clearer dataset/augmentation details will boost confidence in real-world AI performance.
Note in limitations that use beyond Cyprus may need more testing.Models trained on data from one specific geographic region may not perform as well when applied to images from other regions.
Explicitly add app speed wrt to the avg prediction time.
Good study on an important topic. The "I-Tick" web app is a valuable outcome. Sharing code/data is valuable for further research and reproducibility.
Minor improvements including clarifying dataset and experimental methods. This will make a strong paper even better. Well-structured with thoughtful discussion.
Clear and technically sound English.
Some terminologies could be handled better ex. class names like “non-tick”
AI/IoT scope fits journal.
Training setup is well documented.
Results support conclusions.
Need statistical significance testing.
Not sufficient and not so strong error analysis.
This article is highly significant for 2 reasons:
(1) It applies deep learning to a real-world, biologically and medically relevant problem , and
(2) Contribution in vector surveillance and outbreak preparedness using a scalable AI + IoT framework.
Thank you for such a nice article. I enjoyed reading it.
Please refer below for peer review comments.
Abstract:
1. The abstract notes 19,345 images, but it lacks context about how diverse or representative this dataset is. Can you include that information.
2. The abstract lists 7 performance metrics. Instead can you focus on main 3 or 4 and mention others in main results section of this study.
3. You mention “All the 3 models were employed for the development….” – this could confuse readers about which model was chosen for deployment. Please revise this sentence. Consider revising the best performing model.
4. What is clinical or practical implication of this study?
Introduction:
1. Can you explicitly state why current tick identification methods fail. What are inefficiencies in manual classification?
2. I don’t see AI vector tracking efforts beyond tick classification, can you refer any study , its significance and limitations in this area?
3. Phrases like “non-tick images” and “unidentified classes” appear abruptly without any prior context.
Methodology:
1. Can you consolidate all training parameters into one clear sub-section ex. 'Model Training Configuration' , right now they are all distributed across different sections.
2. For custom 5-layer CNN – can you add information regarding each layer’s type, kernel size, activation, and output shape.
3. Can you clarify if any data augmentation techniques were used?
4. When you say overfitting with 80:20 splits , then switched to 50:30:20, this decision needs a clearer statistical or empirical backing. Can you please add that?
5. Spiders and objects are part of “non-tick” images - how these images help improve model performance? Do they introduce any bias in binary/ternary classification?
Results:
1. The use of 7+ metrics is exhaustive. Can you group related metrics (ex. precision/recall/F1) together and summarize them in tables rather than redundant text.
2. Can you add p-values or statistical tests (ex. ANOVA or t-tests) comparing the three models’ results? This would validate that differences are not due to chance.
3. Where did misclassifications occur? You have counts but no qualitative error discussion. Is it between classes?
4. You mention “100% AUC” for all models - This is unusually high and may indicate overfitting. This needs critical discussion. You can have ROC plots and discuss their shape and threshold behavior.
5. Was the performance difference significant for 20 vs. 30 epochs. Can you have this as standalone insight and try to summarize this impact visually ex. use bar chart.
Conclusion:
1. The statement “demonstrated an easier and faster approach of tick identification” must be accompanied with qualified metrics and discussion.
2. For future works you mention “explore other DL models,” – can you specify which architectures ex. DenseNet, EfficientNet, Transformer-based vision models and what challenges they may solve.
3. Do you think human in the loop clarification is needed? Does the system require or benefit from expert review? Please provide your insights. Can you add some guidance on how experts might interact with the model.
I have read over this manuscript and have several comments. First, I am not an expert on AI and automated image identification. My expertise is on the biology of ticks and most of my comments are on tick-specific issues. I was involved in an earlier CNN approach to tick identification but primarily as an expert on ticks. I cannot provide any critical comments on the computational approaches reported here.
There were several cases of questionable or unclear use of terminology and wording related to ticks, and of overly simplistic categorization approaches.
- Separate from identifying genera, most ticks exhibit a 4-stage life cycle where eggs hatch into larvae, larvae develop into nymphs (after a blood meal), and nymphs develop into adults (after a blood meal). There are also male and female adult ticks, as well as engorged vs. flat ticks. It was not clear if or how these complexities are addressed or analyzed (if they are).
- Second, the author’s approach is to distinguish between two genera but in many regions there are four or five or more tick genera. Distinguishing two genera seem like the absolute minimum.
- Third, emphasis is on identifying ticks to the genus and not to species. Are the models capable of this level of discrimination at the species level. Different species within one genus can transmit different pathogens so identification to the species level can have medical significance.
- On line 161 and elsewhere (512), some ticks are referred to as “blacklegged”. There is the black-legged deer tick (Ixodes species) but blacklegged has no scientific name associated with it. Use the scientific name.
- Lines 238-244 uses very casual language for tick traits. It might be better and more accurate to get this information from existing keys or manuals.
In places in the text there are opinions or value judgements that are used as adjectives. For example, line 188 – “would easily surpass any human eye” . Line 194 – “subtle differences that even experts would have no idea about”. Avoid this type of language. Also, line 210 – “models were not hallucinating”.
The design and analyses are beyond my area of expertise, but they seem appropriate.
I would have like to see some discussion about the feasibility of people using these approaches for tick identification. How much work is it to generate the images necessary for the analyses? Is there specialized equipment needed (like cameras, binocular scopes)? Would they be available in a medical office or hospital? Should there be data about pathogens vectored by ticks that are being identified, besides just the genus name?
Line 389 – 407 presents a lot of percentages. Would they be better presented in a figure?
431 – 433 – this seems like a high rate of misidentification.
Somewhere in the introduction or discussion I would have expected to read some text about the benefits and risks (to people receiving tick bites) of automated identification of genera with less than 100% accuracy. What are some of the health benefits or risks to people that go along with this approach? What is the current state of affairs?
Specific comments:
Line 55 – awkward wording
61, 64, 93 – why not species level?
70 – entomology is the science of insects and not mites. Experts in ticks and mites are acharologists.
118 – inconsistent italicizing of tick genera names (they should be italicized).
190 – bugs is not a good work. Arthropods maybe.
259-280 – this seems like a lot of discussion or introduction-type text within the methods section. Similarly, 363-370 seems like introduction/background material in the results section.
282-287 – I do not follow this (out of my area of expertise).
538 - 900 tick species (not genera).
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.