Comparison of Human Intestinal Parasite Ova Segmentation Using Machine Learning and Deep Learning Techniques

Round 1

Reviewer 1 Report

The authors compare the conventional techniques and novel techniques, Deep Learning Techniques which are useful and necessary for clinical diagnosis. However, I have some ethical concerns about human ethics, biosafety, and animal ethics because there are no mentions of the source of helminth ova. To accept for the publication, the authors have to declare the source of the helminth ova in the material and method part that the helminth ova come from the feces of a human patient with consent or laboratory animals. All other data in the article is enough to accept. Only some minor things to improve, especially, the resolution of Figure 1 which is quite low resolution. All helminth ova should have a magnification bar for comparing the size of the ova. 

Author Response

• The ethical concern is added in the “Institutional Review Board Statement” in line 384.
• The source for the helminth ova is added in Section 2.1, lines 109-112. 
• “The Department of Microbiology and Parasitology from Hospital Universiti Sains Malaysia (HUSM) prepared these feces samples, which are freshly obtained from the helminthiasis patients. The feces slides are observed under 40X magnification and saved to JPG extension with a dimension of 1249 × 980.”
• The resolution for Figure 1 is improved. Line 101.
• There is a magnification bar for the ova size during image capturing but has vanished after undergoing a segmentation procedure.
• Please see the attachment for the corrections that have been done.

Author Response File: Author Response.docx

Reviewer 2 Report

The authors have done an excellent job writing a clean and readable manuscript. The present paper has been written coherently. Different sections are organized well and scientifically sound. 

 

However, for further improvements, my suggestions are as follows:

1. Some minor English language grammar correction is needed. I suggest running the manuscript with any correction software (such as Grammarly) or getting it checked with an expert. 

2. From where has the data been collected? There has no source of data been cited by the authors. The readers must be interested to know if the data has been procured from any open source or not. 

3. For reproducibility, in line 427, it would be great if the authors could provide the version of the software ImageJ they used. 

4. Accuracy and IoU have been used as performance metrics. It would be great if the authors shed some light on why they chose these metrics over others. 

5. An appropriate conclusion needs to be written. The authors must also describe any detail of the merit and applicability of such work. Would like tIt would be great if the authors could discuss if there is any room for improvement or further development. 

 

 

 

Author Response

1. Some minor English language grammar correction is needed. I suggest running the manuscript with any correction software (such as Grammarly) or getting it checked with an expert. 

• As suggested, the manuscript has been checked with grammar correction software.

2. From where has the data been collected? There has no source of data been cited by the authors. The readers must be interested to know if the data has been procured from any open source or not. 

•  The source for the helminth ova is added in Section 2.1.

“The Department of Microbiology and Parasitology from Hospital Universiti Sains Malaysia (HUSM) prepared these feces samples, which are freshly obtained from the helminthiasis patients. The feces slides are observed under 40X magnification and saved to JPG extension with a dimension of 1249 × 980.”

3. For reproducibility, in line 427, it would be great if the authors could provide the version of the software ImageJ they used.

• ImageJ software is a freeware tool and has been widely utilized by biologists for its utility and ease of use in handling many sorts of image data across numerous computing systems.”

4. Accuracy and IoU have been used as performance metrics. It would be great if the authors shed some light on why they chose these metrics over others. 

• This is because accuracy and IoU can analyse the quality of the segmented image produced.

5. An appropriate conclusion needs to be written. The authors must also describe any detail of the merit and applicability of such work. It would be great if the authors could discuss if there is any room for improvement or further development.

• Paragraph for the conclusion is added in Section 4. The future work recommendation has also been included in Section 4.

“ The work presented in this paper compared the segmentation performance of machine learning and deep learning when applied to the four helminth ova species based on human intestinal parasite ova which are ALO, EVO, HWO, and TTO. Through the segmentation performance achieved, the deep learning segmentation procedure is more suitable to segment the helminth ova compared to the machine learning segmentation. However, there are still some flaws that should be improved so that the segmentation performance can have a better value.

This work serves as an initial step of image segmentation in developing an automatic detection system for human intestinal parasite ova, hence details can be carried out in the future. Another technique also might be more suitable to ease the segmentation procedure other than the color model and SLIC technique. Then, more comprehensive research comparing other popular segmentation algorithms might be a suitable alternative study to evaluate segmentation performance”

Reviewer 3 Report

Lim et. al.

Applied Sciences

 

Comparison of Human Intestinal Parasite Ova Segmentation using Conventional and Deep Learning Techniques

 

Summary:

In this manuscript, the authors have compared conventional and deep learning segmentation of helminth ovum in light micrographs with an objective toward diagnostic accuracy, accessibility, and throughput for the most affected populations. 

 

Major comments:

This study is nicely executed and may facilitate better penetrance of diagnostic care in sparse, remote, or underserved populations. However, the objective and logic are poorly presented in this manuscript, the analysis is incomplete, and there is no conclusion to the work. These major issues are all fixable, and in the spirit of moving this work forward I have listed the points that must be addressed below and hope to see this manuscript again. Additionally, the English grammar is incredibly rough. The authors have clearly tried, and I thank you all for that, but it does not meet standard and in some places the poor grammar makes the intention unclear. After making the listed changes, I strongly encourage the authors to seek the help of a professional editing service or native English speaker.

 

Lines 318-329 and table 3: The finding is that deep learning produced marginally better segmentation, accuracy being effectively unchanged while the interaction-over-union is highly variable. The IoU variation is interesting and should be analyzed further. The authors need to complete the analysis by investigating the failure rate and modes associated with the IoU; adding a paragraph would be a good idea. The reported percents are an average over a certain number of ovum segmentations in each species and obscure the failure mode(s) in each method for each species. Report the number in each class (n). Show a distribution of segment IoU values for each method in each species to distinguish 100% IoU, 75% of the time from 75% IoU, 100% of the time. Consider whether some ovum shapes are more likely to be mis-segmented and how and why.

 

Line 329: There is no concluding paragraph. Please add a paragraph that connects the findings of this study to the objective.

 

Line 27: Missing statement of that this finding means in the context of the objective.

 

Lines 55-66: This is the description of an unmet medical need and there are hints, but no clear narrative of the part of this problem that THIS study solves. I hope the authors meant to aid efficient delivery of accurate diagnoses to affected populations who often reside in the poorest or most deprived communities (per reference 1). The authors need to clarify whether their findings are addressing diagnostic accuracy, or deliverability to medically underserved communities, or deliverability to medically undersupplied communities, or something else, or some combination thereof. 

Line 60: unreferenced pronoun is unclear what reduces diagnostic accuracy. If the authors mean conventional microscopy by a trained technician, then state a metric of misdiagnosis. This reason is similarly unclear since the accuracy of the reported Conventional method in table 3 is high and similar to the deep learning methods assayed. If this is a misinterpretation of multiple uses of the term conventional, the authors should consider carefully the terms in which they present their story.

Line 64: If rapid response in the field is the objective, then the findings should support a mechanism by which these diagnostics can be put within reach of the care providers who need them. e.g. I can do a fecal egg count on a farm in 15 min with a cheap scale and a children's toy microscope, but there's only one of me and I'm busy. If care providers can upload micrographs for AI analysis to facilitate point of care diagnosis, then this testing might reach the affected patients in a relevant timeframe. Please clarify the objective that this study furthers.

 

Line 90-92: Please state the actual findings of this study here instead of the results-are-achieved nothing that was put here. It should be an improvement over a conventional method.

 

Minor comments:

Line 17: observation on a computer monitor versus microscope objective is not the objective of the present study.

Line 18: an aim to analyze is not a finding. Please report the finding.

Line 34: 'The global report' is inaccurate. This information comes from the WHO global report on soil-transmitted helminth infections.

Line 40: 'mental, and emotional well-being' is missing a reference.

Line 41: 436 million is not 42% of 1.5 M cases in the world, please clarify.

Author Response

 Please see the attachment.

Author Response File: Author Response.docx

Reviewer 4 Report

English language in this paper can be improved. The authors are encouraged to seek help from professional editing services.

Line 108-109, It is understandable that underexposed images can be used to verify performance of contrast enhancement algorithms. I am confused about over-exposed illumination, is the author referring to images with higher brightness without saturating the camera? Do these 66 over-exposed, underexposed images overlap with 100 normal illumination images except lighting conditions? If so, there might be a potential data leak issue when splitting data in model performance assessment

Line 237. “83%. 83%” ? what is it? The author may want to elaborate on what training parameters were used, epochs, batch size, stopping criteria. Etc.

Line 242. Keeping 10% data for validation could be inadequate to properly assess model performance.  By referring to validation, are you talking about validation during training or after training? Ideally, there should be a holdout set to properly assess model performance.

Line 277   why using SLIC before apply CNN segmentation when CNN semantic segmentation usually do a good job alone?

Line 321 it would be nice if there is a more thorough examination as to which cases failed more often for certain segmentation algorithms with some photo examples.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 3 Report

Lim et. al.

Applied Sciences

 

Comparison of Human Intestinal Parasite Ova Segmentation using Machine Learning and Deep Learning Techniques

 

Summary:

In this manuscript, the authors have compared conventional and deep learning segmentation of helminth ovum in light micrographs with an objective toward diagnostic accuracy, accessibility, and throughput for the most affected populations. 

 

Major comments:

The authors have addressed all scientific content issues satisfactorily. 

 

Rather than involving a native English speaker or professional editing service, as previously strongly recommended, the authors have chosen to use grammar correction software. The results are improved, but still sub-par in grammar, punctuation, conjugation, article use, and word choice. I leave it to the editor to decide whether the journal will serve as a copy-editing service.