In-Line Detection of Clinical Mastitis by Identifying Clots in Milk Using Images and a Neural Network Approach

Round 1

Reviewer 1 Report

A network for detecting clots in this study is proposed to improve clinical bovine mastitis. This is an interesting topic, and the research results are very good. If it can be integrated with automatic milking system in the future, it will improve the sensitivity of clinical mastitis detection during automatic milking. However, there are several issues which need to be fixed properly.

1. The author should provide theoretical proof (references) in defining the main theme in first paragraph of introduction.

2. Author should improve defining the consistency control problem. Description given problem is not enough is just presented models proposed time and name...rather it should provide the problem parameters which effect consistency.

3. The authors used the flash of the iPhone 6 to take photos while obtaining data. The data of 1282 photos is a little small, and you can consider the use of non-flash testing when testing. Can the same achieve 100% training effect？

4. There is no "deep learning" or "image recognition" in the abstract, but in my opinion, it is replaced with "neural network" and " identified the clots ".

Extensive editing of English language required

Author Response

1. The author should provide theoretical proof (references) in defining the main theme in the first paragraph of introduction.

Thank you for reviewing our article. We have added additional references to our introduction.

2. Author should improve defining the consistency control problem. Description given problem is not enough is just presented models proposed time and name...rather it should provide the problem parameters which effect consistency.

We appreciate your suggestion to elaborate on consistency control. To address this, we have defined a clearer methodology (Tables 1 and 2) in order to increase the reproducibility of our results.

3. The authors used the flash of the iPhone 6 to take photos while obtaining data. The data of 1282 photos is a little small, and you can consider the use of non-flash testing when testing. Can the same achieve 100% training effect？

We acknowledge the point raised concerning our dataset size and the use of flash in image capture. We will explore the possibility of augmenting our dataset with non-flash images to validate the robustness of our model against varying lighting conditions, although the current dataset has confidently supported our findings of a pilot test of the technology. We have added this to the discussion:

The current study encountered limitations due to the relatively small dataset, which lacked diversity. Although image augmentation techniques were applied to introduce variability, such approaches do not compare to larger, more complex datasets typically utilized in deep learning studies [30]. Furthermore, the consistency of the recording setup throughout the study, exclusively using an iPhone's flash for illumination, has left the model's robustness to alternate lighting conditions untested, a notable concern since, in field conditions, lighting can be inconsistent and obstructed. Consequently, the findings presented here should be interpreted as preliminary, serving as an exploratory investigation into the application of deep learning for mastitis detection. It is recommended that future research be conducted with more extensive datasets gathered from field conditions on AMS to thoroughly evaluate the model's performance and practicality.

4. There is no "deep learning" or "image recognition" in the abstract, but in my opinion, it is replaced with "neural network" and " identified the clots ".

Your observation about the absence of the terms 'deep learning' and 'image recognition' in the abstract is valid. We have revised the abstract (lines 23 and 33) to explicitly include these terms, as they are integral to the methodology and thus should be highlighted appropriately.

Reviewer 2 Report

This is potentially a very valuable paper that describes a NN approach to the detection of mastitis in automated milking systems. The introduction, materials and methods section are very good.  However, the 4-line description of your results that claims 'The accuracy.....and sensitivity......... were 100%'. (take our word for it) will not do. You need to provide at least a Table giving numbers of tests and figures for accuracy, sensitivity etc.

Author Response

Thank you for your positive appraisal of our article. Indeed, the numbers are quite vague by just saying 100%. We’ve added an additional confusion matrix (table 3) with the actual number of figures used in the holdout testing group.