Review of Quantitative Methods for the Detection of Alzheimer’s Disease with Positron Emission Tomography
Round 1
Reviewer 1 Report
Review of Quantitative Methods for the Detection of Alzheimer’s Disease with Positron Emission Tomography
A review is provided of the data-driven methods available for molecular neuroimaging studies, namely FDG, amyloid, tau, pet, with an emphasis on the use of machine/deep learning as quantitative tools complementing the specialist in detecting AD.
The article delivers a sufficient review, it is however missing some critical points and details. The language should be made more concise. Points are detailed below with reference to sentence number on the draft document.
#1 You jump straight into background, you need an introduction laying out the review and the sections of the paper.
Figure 1 is not referenced in the text
Reference the prevalence of AD, global absolute figures are difficult to comprehend
sentence numbers referenced*
54-58: very long sentence, break into two.
70: communication of multidomain ....AD patients
73: in CSF; tau. You use lots of commas preceding "and" which makes difficult reading, use semicolons or break the sentence up.
111: Ref[10] doesn't relate to quantitative methods having great potential but not used clinically
129: mention oxygeneation
e.g Fan, A. P., An, H., Moradi, F., Rosenberg, J., Ishii, Y., Nariai, T., ... & Zaharchuk, G. (2020). Quantification of brain oxygen extraction and metabolism with [15O]-gas PET: a technical review in the era of PET/MRI. Neuroimage, 220, 117136.
130: reword - currently the clinical norm, or approaching the
#157 figure 1, colours on figure not defined, what do they represent?
230: paragraph, reference to accuracy of Tau Vs Amyloid, using [47] Ossenkoppele, R et al
#3 You should include the utility of dynamic PET and state static PET being semi quantitative.
242-249: I understand the papers focus is on machine learning but an introduction to standard quantitative imaging (e.g SUVr) for AD is warranted including its limitations and comments around standardization around this, also pitfalls wrt. acqusistion variabilities (e.g partial volume corrections, acquisition time, recon, effect quantification)
You should introduce "why" we need (advanced) data driven methods.
You have also not mentioned radiomics in the more detailed section a reference to this would be pertinent e.g [1]
[1]Li, L., Yu, X., Sheng, C., Jiang, X., Zhang, Q., Han, Y., & Jiang, J. (2022). A review of brain imaging biomarker genomics in Alzheimer’s disease: implementation and perspectives. Translational neurodegeneration, 11(1), 1-37.
#256 no big M defined in eq 1
#257 "features proper" incorrect language
#262 The derivation of OLS is excessive detail if you need it just put matrix notation for eq 1 and eq3 (as you have done for PCA)
#262 equation 2 you need subscripts on x and e, if your not using matrix notation
#264 equation 3 isn't necessary, just state minimizing the square of the sum of the residuals is satisfied by ....B
#342 typo St-Aubert
#276 figure text is unclear, figure 2 not referenced in the text
#include the disadvantages of SPM
#426 software package, this has been (grammar)
#438, reword, remove obscure (in accurate / grammar)
#472 highly engineering (grammar)
#479 optimal re-express (optimally), data is (not are), poorly worderd, optimal in what sense? Mention maximizing variance, minimizing reconstruction error
#487 data is also normalized, unit variance, please include
#488 ijth component not required to much detail
#493 The covariance of matrix (re-word)
#503 "in the end" The PCA section needs rewording, place 493-495 nearer 476, you introduce SVD in 532, include this in the diagnolization of PCA. Start with PCAs motivation and result of how you get there, projecting with eigenvectors then follow up with "some" maths, you don't need a full derivation, mention eigenimages and how they have been used in the field of AD research.
#505 Put section relating SPM / SSM / PCA e.g [2]
[2] Hocurscak, L., Tomanic, T., Trost, M., & Simoncic, U. (2021). Comparison of statistical parametric mapping method and scaled subprofile model for functional neuroimage analysis. In APS March Meeting Abstracts (Vol. 2021, pp. F15-002).
#550 you haven't referenced fig 3 in the text
#626 "We will now explore the use of machine learning (ML) in the context of the support vector machine" PCA is also machine learning, the way you have introduced machine learning suggests not, please adress.
#Include discussion on supervised (SVM) and unsupervised (PCA) learning, include PCA also being used as a pre-processing step for further analysis, also include the use of PCA regression.
#665 you haven't referenced fig 5 in text
#703 ADNI acronym not introduced
#SUVR not defined
#737 – the perpetual, please be clearer this is not accurate use of language
#738 Deep learning (DL) methods have the automatic advantage over ML .. in that they do not rely on any a priori knowledge, neither does unsupervised learning, be clearer. This phrase has many holes, you also need to be clear in that you are making a distingushing difference between DL and ML (DL is a subset of machine learning with the advantage of ... )
#746 instead use their prior knowledge ( I thought they didn't need priori knowledge?). This section needs to be more consice, it is repetitive in parts. Introduce neural networks first, then deep learning
#767 "would be collapsible" grammar
Fig 6, fig 7, not referenced in text
You should mention voxelwise learning before introducing inception V3 studies. Also the utility of voxel wise learning (e.g [3]). You go on to reference studies which use image generation so this needs to be made clear, your figures (6,7) suggest networks are just used for classification.
[3] Ruwanpathirana, G. P., Williams, R. C., Masters, C. L., Rowe, C. C., Johnston, L. A., & Davey, C. E. (2022). Mapping the association between tau-PET and Aβ-amyloid-PET using deep learning. Scientific reports, 12(1), 1-11.
#823 "gaining recognition" is vague, please include. GANs, are a type of generative model that learn the distribution of training images and have been utilized for PET AD image classification [4]. The common theme in GANs and RNNS is the basis of .....
[4]Penning, J., John, R., Chandler, H., Fielding, P., Marshall, C., & Smith, R. (2021, October). Generative Adversarial Network" Steerability" for Brain PET Image Generation. In 2021 IEEE Nuclear Science Symposium and Medical Imaging Conference (NSS/MIC) (pp. 1-4). IEEE.
#857 Refernce [133] doesn't have a journal listed, I also cant find this online, please update or include a new references
[133] Ryoo, H. G., Choi, H., & Lee, D. S. (2022). Distinct subtypes of spatial brain metabolism patterns in Alzheimer’s disease identified by deep learning based FDG PET clusters.
Please include variational autoencoders have however demonstrated varying accuracy dependent on the quality of the training data [5].
[5]John, R., Penning, J., Chandler, H., Fielding, P., Marshall, C., & Smith, R. (2021, October). Quantitative Evaluation of Synthesized Brain PET Using a Variational Autoencoder. In 2021 IEEE Nuclear Science Symposium and Medical Imaging Conference (NSS/MIC) (pp. 1-4). IEEE.
#901 please include reference to lack of transparency
#922 autoencoders that can reconstruct any transformed data, please reword as reconstruction isn't always faithful (depends on the training data)
Author Response
Point 1: You jump straight into background, you need an introduction laying out the review and the sections of the paper.
Response 1: An introduction section has been added to include clinical standard for dementia evaluation, motivation for the use of data-driven methods, and motivation for PET in particular.
Point 2: Figure 1 is not referenced in the text
Response 2: This has been corrected.
Point 3: Reference the prevalence of AD, global absolute figures are difficult to comprehend
Response 3: A prevalence has been added for ease of understanding.
Point 4: 54-58: very long sentence, break into two.
Response 5: you are correct, this has been corrected
Point 6: 70: communication of multidomain ....AD patients
Response 6: typographical error has been corrected
Point 7: 73: in CSF; tau. You use lots of commas preceding "and" which makes difficult reading, use semicolons or break the sentence up.
Response 7: I have eliminated most Oxford commas and added sentences to reduce the difficulty of reading
Point 8: 111: Ref[10] doesn't relate to quantitative methods having great potential but not used clinically
Response 8: You are correct. This mistake has been fixed with an appropriate reference.
Point 9: 129: mention oxygeneation
Response 9: Reference and mention of O-15 PET has been added for completeness using the reference kindly provided
e.g Fan, A. P., An, H., Moradi, F., Rosenberg, J., Ishii, Y., Nariai, T., ... & Zaharchuk, G. (2020). Quantification of brain oxygen extraction and metabolism with [15O]-gas PET: a technical review in the era of PET/MRI. Neuroimage, 220, 117136.
Point 10: 130: reword - currently the clinical norm, or approaching the
Response 10: This section has been reworded.
Point 11: #157 figure 1, colours on figure not defined, what do they represent?
Response 11: Each color represent a distinct ROI in the BrainViewer atlas. This has been added in the figure caption for clarity.
Point 12: 230: paragraph, reference to accuracy of Tau Vs Amyloid, using [47] Ossenkoppele, R et al
Response 12: I have added the appropriate reference and discussed tau-PET outperforming amyloid-PET and volumetric MRI
Point 13: #3 You should include the utility of dynamic PET and state static PET being semi quantitative.
Response 13: a paragraph on the usage of dynamic and static PET has been added in the context of introducing kinetic modelling and semi-quantitative methods
Point 14: 242-249: I understand the papers focus is on machine learning but an introduction to standard quantitative imaging (e.g SUVr) for AD is warranted including its limitations and comments around standardization around this, also pitfalls wrt. acqusistion variabilities (e.g partial volume corrections, acquisition time, recon, effect quantification)
Response 14: significant detail regarding these concepts have been added per your advice
Point 15: You should introduce "why" we need (advanced) data driven methods.
Response 15: motivation for data-driven methods has been added to both the new introductory section and the section on methodologies
Point 16: You have also not mentioned radiomics in the more detailed section a reference to this would be pertinent e.g [1]
[1]Li, L., Yu, X., Sheng, C., Jiang, X., Zhang, Q., Han, Y., & Jiang, J. (2022). A review of brain imaging biomarker genomics in Alzheimer’s disease: implementation and perspectives. Translational neurodegeneration, 11(1), 1-37.
Response 16: This is a great reference. It has been added to the appropriate section and a brief mention of radiomics has been added
Point 17: #256 no big M defined in eq 1
Response 17: Error corrected
Point 18: #257 "features proper" incorrect language
Response 18: error corrected
Point 19: #262 The derivation of OLS is excessive detail if you need it just put matrix notation for eq 1 and eq3 (as you have done for PCA)
Response 19: the derivation of OLS has been reduced and simplified
Point 20: #262 equation 2 you need subscripts on x and e, if your not using matrix notation
Response 20: I believe the use of matrix notation is clearly implied
Point 21: #264 equation 3 isn't necessary, just state minimizing the square of the sum of the residuals is satisfied by ....B
Response 21: this has been eliminated and streamlined (see point 19)
Point 22: #342 typo St-Aubert
Response 22: typo (additional space) corrected
Point 23: #276 figure text is unclear, figure 2 not referenced in the text
Response 23: figure has been simplified and mentioned in the text
Point 24: include the disadvantages of SPM
Response 24: several notable disadvantages of SPM has been included
Point 25: #426 software package, this has been (grammar)
Response 25: grammar corrected
Point 26: #438, reword, remove obscure (in accurate / grammar)
Response 26: inaccurate statement has been removed
Point 27: #472 highly engineering (grammar)
Response 27: incorrect grammar corrected
Point 28: #479 optimal re-express (optimally), data is (not are), poorly worderd, optimal in what sense? Mention maximizing variance, minimizing reconstruction error
Response 28: data is plural of datum and should not be used as a singular noun, grammatical error "optimal re-express" has been corrected, mention of maximizing variance and minimizing reconstruction error added
Point 29: #487 data is also normalized, unit variance, please include
Response 29: included
Point 30: #488 ijth component not required to much detail
Response 30: this phrase has been removed
Point 31: #493 The covariance of matrix (re-word)
Response 31: this phrase has been reworded
Point 32: #503 "in the end" The PCA section needs rewording, place 493-495 nearer 476, you introduce SVD in 532, include this in the diagnolization of PCA. Start with PCAs motivation and result of how you get there, projecting with eigenvectors then follow up with "some" maths, you don't need a full derivation, mention eigenimages and how they have been used in the field of AD research.
Response 32: relevant section has been reworded and restructured, some mathematics have been eliminated, eigenimages/eigenbrains have been integrated into the section
Point 33: #505 Put section relating SPM / SSM / PCA e.g [2]
[2] Hocurscak, L., Tomanic, T., Trost, M., & Simoncic, U. (2021). Comparison of statistical parametric mapping method and scaled subprofile model for functional neuroimage analysis. In APS March Meeting Abstracts (Vol. 2021, pp. F15-002).
Response 33: appropriate citation and mention of analytical relationship between SPM and SSM/PCA have been added
Point 34: #550 you haven't referenced fig 3 in the text
Response 34: this has been corrected
Point 35: #626 "We will now explore the use of machine learning (ML) in the context of the support vector machine" PCA is also machine learning, the way you have introduced machine learning suggests not, please adress.
Response: you are correct and language has been cleaned up to be more specific
Point 36: #Include discussion on supervised (SVM) and unsupervised (PCA) learning, include PCA also being used as a pre-processing step for further analysis, also include the use of PCA regression.
Response 36: details about PCA has a form of machine learning (unsupervised) and regression PCA have been added, please note the multiple mentions of PCA/SSM as "preprocessing" for SVM-based classification studies in that section
Point 37 #665 you haven't referenced fig 5 in text
Response 37: this has been corrected
Point 38: #703 ADNI acronym not introduced
Response 38: this has been corrected
Point 39: #SUVR not defined
Response 39: SUVR added in section on traditional quantitative PET modeling and explicitly defined where original flagged by reviewer
Point 40: #737 – the perpetual, please be clearer this is not accurate use of language
Response 40: this has been changed and removed to be more precise
Point 41: #738 Deep learning (DL) methods have the automatic advantage over ML .. in that they do not rely on any a priori knowledge, neither does unsupervised learning, be clearer. This phrase has many holes, you also need to be clear in that you are making a distingushing difference between DL and ML (DL is a subset of machine learning with the advantage of ... )
Response 41: these misleading phrase have been removed and replaced with specific discussion of neural networks
Point 42: #746 instead use their prior knowledge ( I thought they didn't need priori knowledge?). This section needs to be more consice, it is repetitive in parts. Introduce neural networks first, then deep learning
Response 42: the section on NN has been edited to be somewhat more concise and misleading language has been eliminated
Point 43: #767 "would be collapsible" grammar
Response 43: relevant statement has been removed to make section more concise
Point 44: Fig 6, fig 7, not referenced in text
Response 44: this has been corrected
Point 45: You should mention voxelwise learning before introducing inception V3 studies. Also the utility of voxel wise learning (e.g [3]). You go on to reference studies which use image generation so this needs to be made clear, your figures (6,7) suggest networks are just used for classification.
[3] Ruwanpathirana, G. P., Williams, R. C., Masters, C. L., Rowe, C. C., Johnston, L. A., & Davey, C. E. (2022). Mapping the association between tau-PET and Aβ-amyloid-PET using deep learning. Scientific reports, 12(1), 1-11.
Response 45: this reference to the relative performance of CNNs and GLMs has been added
Point 46: #823 "gaining recognition" is vague, please include. GANs, are a type of generative model that learn the distribution of training images and have been utilized for PET AD image classification [4]. The common theme in GANs and RNNS is the basis of .....
[4]Penning, J., John, R., Chandler, H., Fielding, P., Marshall, C., & Smith, R. (2021, October). Generative Adversarial Network" Steerability" for Brain PET Image Generation. In 2021 IEEE Nuclear Science Symposium and Medical Imaging Conference (NSS/MIC) (pp. 1-4). IEEE.
Response 46: this is also an excellent reference, appropriate changes to phrasing have been made
Point 47: #857 Refernce [133] doesn't have a journal listed, I also cant find this online, please update or include a new references
[133] Ryoo, H. G., Choi, H., & Lee, D. S. (2022). Distinct subtypes of spatial brain metabolism patterns in Alzheimer’s disease identified by deep learning based FDG PET clusters.
Response 47: appropriate reference has been updated
Point 48: Please include variational autoencoders have however demonstrated varying accuracy dependent on the quality of the training data [5].
[5]John, R., Penning, J., Chandler, H., Fielding, P., Marshall, C., & Smith, R. (2021, October). Quantitative Evaluation of Synthesized Brain PET Using a Variational Autoencoder. In 2021 IEEE Nuclear Science Symposium and Medical Imaging Conference (NSS/MIC) (pp. 1-4). IEEE.
Response 48: this has been amended to reflect that the quality of training data affects accuracy of reconstruction
Point 49: #901 please include reference to lack of transparency
Response 49: reference has been added
Point 50: #922 autoencoders that can reconstruct any transformed data, please reword as reconstruction isn't always faithful (depends on the training data)
Response 50: this has been added, please see point 48
Reviewer 2 Report
The authors proposes an interesting review of quantitative methods for Alzheimer detection using PET.
Although the paper is very rich in informations I find it hard to read.
First it does not follow the standard structure as there is no introduction and no conclusion.
The choice and structuring of the sections is vague and unjustified. The authors should first explain on what basis they will classify the different approaches.
It would be interesting to present the synthesis of the discussion, and possibly of certain sections. In the form of a table for example, to better present the highlight of the review.
Author Response
Point 1: Although the paper is very rich in informations I find it hard to read.
Response 1: There has been extensive rewriting to make the paper easier to read and simplify the language chosen
Point 2: First it does not follow the standard structure as there is no introduction and no conclusion.
Response 2: We have added these sections.
Point 3: The choice and structuring of the sections is vague and unjustified. The authors should first explain on what basis they will classify the different approaches.
Response 3: We have added text at the beginning of the section describing these methods which details how these methods have been grouped/classified.
Point 4: It would be interesting to present the synthesis of the discussion, and possibly of certain sections. In the form of a table for example, to better present the highlight of the review.
Response 4: A table has been added at the end of the section describing the quantitative methods comparing performance of various methods for the detection of AD/MCI pathologies for ease of comparison.
Round 2
Reviewer 1 Report
All previous issues addressed, there are a few typos, one "[ref]" where there is a missing reference, address these, but I accept.
Reviewer 2 Report
The article has significantly improved in this latest version. My main remarks have been corrected.
I recommend the publication
