Datasets:

mfromm
/

AMSR

{"forum": "BJgm2DkJgN", "submission_url": "https://openreview.net/forum?id=BJgm2DkJgN", "submission_content": {"title": "Spherical CNN-Based Brain Tissue Classification Using Diffusion MRI", "authors": ["Enes Albay", "Gozde Unal"], "authorids": ["albay@itu.edu.tr", "gozde.unal@itu.edu.tr"], "keywords": ["Brain tissue classification", "diffusion MRI (dMRI)", "Spherical CNN (SCNN)", "Fiber Orientation Distribution Function (fODF)"], "TL;DR": "Because fODFs live on the sphere, we propose to label each voxel of diffusion MR image using Spherical CNN (SCNN). ", "abstract": "We propose a method for classification of brain tissue using diffusion MRI data. First, a fiber orientation distribution function (fODF) is constructed at each voxel using Constrained Spherical Deconvolution (CSD) algorithm. Then, instead of secondary properties of reconstructed fODFs, because fODFs live on the sphere, we propose to classify each voxel using Spherical CNN (SCNN) without any transformation into other spaces such as a planar space. Our approach does not require a large number of subjects in contrast to streamline CNN based methods for structural MR image labeling. We present results on a dataset taken from HCP database to demonstrate that our method is suitable to the nature of diffusion data and furthermore it shows transfer capability among subjects.", "pdf": "/pdf/3940f6c8bfe62b964b9404110e62f98ddb1b6cf0.pdf", "code of conduct": "I have read and accept the code of conduct.", "paperhash": "albay|spherical_cnnbased_brain_tissue_classification_using_diffusion_mri"}, "submission_cdate": 1544644523001, "submission_tcdate": 1544644523001, "submission_tmdate": 1545069819051, "submission_ddate": null, "review_id": ["r1gkazwFQ4", "BJxFNWBc7E", "r1xJMipE7N"], "review_url": ["https://openreview.net/forum?id=BJgm2DkJgN&noteId=r1gkazwFQ4", "https://openreview.net/forum?id=BJgm2DkJgN&noteId=BJxFNWBc7E", "https://openreview.net/forum?id=BJgm2DkJgN&noteId=r1xJMipE7N"], "review_cdate": [1548477110802, 1548534065341, 1548176134776], "review_tcdate": [1548477110802, 1548534065341, 1548176134776], "review_tmdate": [1548856747943, 1548856736905, 1548856717783], "review_readers": [["everyone"], ["everyone"], ["everyone"]], "review_writers": [["MIDL.io/2019/Conference/Paper27/AnonReviewer1"], ["MIDL.io/2019/Conference/Paper27/AnonReviewer2"], ["MIDL.io/2019/Conference/Paper27/AnonReviewer3"]], "review_reply_count": [{"replyCount": 0}, {"replyCount": 0}, {"replyCount": 0}], "review_replyto": ["BJgm2DkJgN", "BJgm2DkJgN", "BJgm2DkJgN"], "review_content": [{"pros": "The authors have proposed a SCNN-based brain tissue classification method using DMRI data. The idea of applying SCNN on fODF is straightforward and looks interesting. The paper is well written and it is easy to follow. ", "cons": "But, I have concerns on the evaluation of the proposed method.\n\n-\tThere is no comparison to other methods. To convince applying SCNN is promising, a comparison, as least, to a method with CNN on fODF directly could be helpful. There is one dice score from an existing study reported. Did this study analyze the same data?\n-\tI do not see why it is a good idea to train on a single subject. Including more subjects for training to account for individual variability should be a much better idea.\n-\tThe performance on CSF is very low (around 60 of DC). Discussion about the performance on this could be helpful.\n-\tThe HCP data provides tissue segmentation label map in diffusion space. What is the reason to perform reference labeling again? ", "rating": "2: reject", "confidence": "3: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, {"pros": "- definition of a novel CNN approach on the fODF, by applying convolutions to data that lives on SO(3).", "cons": "- geometric distortions in diffusion data are significantly larger than in traditional T1w and T2w data. I am not aware of any studies that would acquire diffusion data only and then employ that data for structural volume analysis. Thus, the need for tissue segmentation from dMRI is significantly lower and really only necessary for the purpose of masking/seeding in diffusion analyses. \n- diffusion MRI Segmentations are compared to FSL-Fast segmentations on structural MRI. That is NOT a reference segmentation as FSL-Fast can fail, be significantly imperfect at given parts in the images, etc. References should be of manual or semi-manual order. Evaluation data is small (16 subjects)\n- Since fODF segmentations are compared to structural MRI segmentation (where the latter is used as segmentation), the obvious question is why not solve this on the structural MRI\n\n", "rating": "2: reject", "confidence": "3: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, {"pros": "\nSUMMARY\n\nThe paper \u201cSpherical CNN-Based Brain Tissue Classification Using Diffusion MRI\u201d presents a neural network that classifies reconstructed diffusion weighted MRI signals into white matter, gray matter and cerebrospinal fluid. \nThe network utilizes spherical convolutional layers (sCNN) with rectified linear units (ReLU) as activation function and ends with fully-connected layers that perform the classification task. Training is based on constrained spherical deconvolution (CSD) orientation distribution functions (fODF) as input and anatomical FAST segmentations as label of a single (human connectome project) subject.\nEvaluation is performed in an inter- and intra-subject manner within the HCP project.\n\nPROS\n \nThe proposed approach utilizes a new and - for the field of diffusion imaging - very interesting method: the spherical convolution.\n\n", "cons": "\nCONS\n\nUnfortunately, there are numerous weaknesses, hence only the most serious ones will be covered here:\n\n(1) The chosen network input:\nIn order to find a good response function (RF) for a CSD reconstruction, a meaningful white matter mask is required for big datasets due to computational constraints. Therefore, using CSD fODFs as input to predict the white matter mask, which is required during generation of the input signal, does not make much sense. \nFurthermore, it should be taken into account that the fODF was generated by deconvoluting the diffusion signal with a single RF. It should therefore be easily possible for a network to learn a convolution, while the plain diffusion signal can be utilized as input.\n\n(2) The networks structure:\nMain purpose of the sCNN layers is to keep the spherical signal structure from layer to layer.\nSince the goal is to classify the input, keeping the spherical structure does not seem important for a good classification. \nFurthermore, applying ReLUs to the Spherical Harmonic signal completely removes this spherical structure, since all values <0 are set to 0. Applying different activation function (e.g. sigmoid or tanh) would most probably keep the spherical structure, in case it might be beneficial for classification. \n\n(3) Evaluation:\nThe biggest drawback of the current evaluation is that no other method was evaluated for comparison. The easiest way to compute a segmentation would be to apply FSL\u2019s FAST on the b=0 diffusion weighted signal. Another possible comparison would be a four-layer neural network with 16, 32, 128 and 3 neurons per layer. This would prove the possible improvement due to the spherical structure.\nThe statement that the network can also be applied to other datasets/subjects needs further investigation, since the HCP Project is a very homogeneous dataset. To this end, it would have been important to evaluate other scanners, different resolutions and different numbers of gradient directions. For a proper evaluation, at least the resolution and the number of gradient directions should be evaluated, as these have a direct influence on the fODF.\n\nCONCLUSION\nThis paper utilizes an interesting network structure for an important task within the field of diffusion imaging. Unfortunately, it doesn\u2019t get far with it.\nAs the paper states itself, only preliminary results are presented. It would therefore be recommended to further improve this work.\n\n\n", "rating": "1: strong reject", "confidence": "3: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}], "comment_id": ["rJl0FdnH4N", "Hkxl-F2BEE", "Bygzc23B4V", "rJlvLKnS4N", "SkgdWDbUNN", "S1l3oi7P4V", "BJe7LTVd4N"], "comment_cdate": [1549285509937, 1549285623864, 1549286537761, 1549285711148, 1549305599715, 1549380516447, 1549450570756], "comment_tcdate": [1549285509937, 1549285623864, 1549286537761, 1549285711148, 1549305599715, 1549380516447, 1549450570756], "comment_tmdate": [1555946035867, 1555946035656, 1555946035442, 1555946035223, 1555946032477, 1555946029735, 1555946026204], "comment_readers": [["everyone"], ["everyone"], ["everyone"], ["everyone"], ["everyone"], ["everyone"], ["everyone"]], "comment_writers": [["MIDL.io/2019/Conference/Paper27/Authors", "MIDL.io/2019/Conference"], ["MIDL.io/2019/Conference/Paper27/Authors", "MIDL.io/2019/Conference"], ["MIDL.io/2019/Conference/Paper27/Authors", "MIDL.io/2019/Conference"], ["MIDL.io/2019/Conference/Paper27/Authors", "MIDL.io/2019/Conference"], ["MIDL.io/2019/Conference/Paper27/Area_Chair1", "MIDL.io/2019/Conference"], ["MIDL.io/2019/Conference/Paper27/AnonReviewer1", "MIDL.io/2019/Conference"], ["MIDL.io/2019/Conference/Paper27/AnonReviewer3", "MIDL.io/2019/Conference"]], "comment_reply_content": [{"replyCount": 0}, {"replyCount": 0}, {"replyCount": 0}, {"replyCount": 0}, {"replyCount": 0}, {"replyCount": 0}, {"replyCount": 0}], "comment_content": [{"title": "Purpose of the paper is show that it is possible to achieve brain tissue segmentation using only dMRI data where the segmentation is difficult with T1w and T2w modalities", "comment": "Dear reviewer 2:\n We thank the reviewer for her/his comments.\n- The purpose of the paper is show that it is possible to achieve brain tissue segmentation using only dMRI data. As mentioned at the conclusion part of the paper, this approach can be used in a multimodal segmentation scenario or for segmentation of infant brains where the segmentation is difficult with T1w and T2w modalities.\n- Because FSL tools may fail at some regions such as around boundaries between gray matter and white matter, we applied a threshold to get rid of those possible FSL-segmentation errors. Certainly, it is possible for us to utilize manual-segmentations, however, the challenge is to obtain a dataset where T1-, T2-weighted images, the manual brain tissue segmentation maps, as well as the diffusion MRI data are all available. FSL segmentations are a current standard tool in our domain. We hope that this explanation clarifies our point to the reviewer. \n- We could not understand exactly what you mean by fODF segmentations, as we do not perform such a task. Our point in this paper is to show that brain tissue segmentation is possible with diffusion MRI data. It would be possible to compare our segmentation results with segmentations over T1- and T2-weighted images if a dataset with manual tissue segmentation was available along with the diffusion MRI data of the same subjects."}, {"title": "fODFs do not live on a Euclidean geometry, it would not be directly applicable to use with CNNs, where planar convolutions are involved", "comment": "Dear reviewer 1,\n We thank the reviewer for her/his comments.\n- As fODFs do not live on a Euclidean geometry, it would not be directly applicable to use with CNNs, where planar convolutions are involved. That is indeed the main idea behind our work. We utilize the SCNN, which works on an appropriate manifold to the fODF structure, i.e. the unit sphere and/or SO(3). There is one dice score because in the literature we found only one study which applied brain segmentation on diffusion MRI data. Our method and results highly outperform that technique.\n- We show in the paper that one of the main strengths of our method is that it can provide good performance results even when we have only limited data. That\u2019s why we showcased a scenario where we train the method only by a single subject and segment other subjects using the single subject trained network. We think that this is definitely not a con, but it is a very important pro side of our proposed method.\n- The performance of CSF segmentation using only diffusion MRI data is also low in the previous method in the literature. This is due to the more unrestricted diffusion of water molecules both in gray matter and csf tissue compared to the white matter. However, the main target for us was to differentiate White Matter and Gray matter, and White Matter and CSF because we are using diffusion MRI data, which can be said to mainly focus on white matter tissue. Our goal in the next steps of our work is to combine other modalities to obtain a multimodal segmentation to address this issue. However, as we stated above, we observed that we achieve better CSF segmentations than the main competing paper on diffusion MRI segmentation.\n- We observed tissue segmentation label problems in the provided HCP labels. By the way, HCP also utilized an automatic tool such as FSL-FAST to perform those segmentations. That\u2019s why we performed FLS-FAST segmentation in order to obtain the tissue segmentation maps.\n"}, {"title": "We are not convinced that all three reviewers have indeed understood the details and contribution of our paper properly. They may start with a huge negative unfair bias against our paper.", "comment": "Dear Chairs,\n\nWe are not convinced that all three reviewers have indeed understood the details and contribution of our paper properly. This leads us to suspect that they were not experienced or did not have an in-depth understanding over the topics of diffusion MRI or the Spherical CNNs. Although they made wrong comments and exhibited wrong knowledge in their feedback which we addressed, all three reviewers indicated they are absolutely certain in their evaluation, and very familiar with the literature.\n\nFurthermore, we would like to point to a flaw we suspect to exist in the OpenReview system. The first reviewer posted their quite negative review with a \"strong reject\" (which is not justified at all) 4 days before the other two reviewers. We wonder whether that review was visible to the other 2 reviewers then? If yes, this certainly affects the other two reviewers, who were not able to return with a positive review, but they started with a huge negative unfair bias against our paper. If this was the case, we ask the Program Chairs and Conference chairs to pay attention to this very important detail in their evaluations and in future versions of MIDL.\n\nIn the end, we would like to ask the Area Chair to take into account our considerations above, and evaluate our paper, the reviews and our rebuttal. We believe it is a worthy contribution, which is very suitable to the MIDL conference.\n\nKind Regards,\nEnes Albay, PhD Candidate\nITU Computer Engineering Department"}, {"title": " The chicken-egg problem which the reviewer claims does not exist in our work", "comment": "Dear reviewer 3,\n We thank the reviewer for her/his comments.\n1) RF function can calculated for each subject in a few minutes, therefore, it does not constitute a big computational constraint. Most importantly, the reviewer has a confusion: in computation of the RF and during CSD, one does not have to utilize a white matter mask. Typically, during RF computation, an FA threshold (e.g. 0.8) is applied and all those voxels with FA higher than threshold are used in the computations. Therefore, the chicken-egg problem which the reviewer alludes DOES NOT EXIST in our work!\nFurthermore, RF function computation is analytically known and there is no need to involve a neural network to make that deconvolution, as we typically utilize neural networks to model functional mappings, which are not analytically defined. Also, as dMRI can be considered to be basically 4D data and learning an RF using a network would require much more computation.\n\n2) Unfortunately, the reviewer is mistaken. Please go ahead and see the \"Spherical CNNs\" paper, after the first layer, the network does not try to maintain spherical structure in each layer, it only receives the initial input as a spherical signal then each layer produces output in SO(3). Therefore, as shown and USED in the original SCNN paper, it is legitimate to use ReLU after each layer safely, which does not distort the SO(3) structure.\n\n3) We would like to point out that the properties of the b0 diffusion weighted signal resembles very much a T2weighted signal in nature. We already utilize both T2 and T1 in the FLS-fast segmentation technique, therefore, we do not believe any gain in the comparison task that the reviewer recommended. The comparison to a standard CNN is already on our to-do-list, but due to space constraints in the paper, we left it for an extension after the MIDL conference. \nCertainly, we are planning to apply our method to other datasets including other scanners and different protocols, however, one should consider the fact that MIDL, like other conferences, has a certain page limit, and it can be considered an initial showcase venue for our work. After presenting the main proof of concept and initial validations over HCP dataset, our aim is to extend our work further. \nWe believe venues such as MIDL, are not the places to showcase the final extended products of research work. On the contrary, it should be a place to present novel ideas with initial evaluations, which we believe our paper presents. For the first time, we showed the plausibility of using diffusion MRI data for brain tissue labeling. Our ongoing work includes its improvements with including other modalities by extending the deep network model with standard CNNs and exploiting datasets in addition to HCP, if they are available.\n\n"}, {"title": "Acknowledged your comment", "comment": "Dear Enes,\n\nWe are very sorry there is a feeling of a possible inexperience from the reviewers. For information, two are Full senior Professors. All are active in the field of diffusion imaging. \n\nWhile this may not guarantee their in-depth understanding of the paper, we have further asked all three reviewers to engage in the discussion. We will, however, let them build their own conclusions. \n\nWe will overlook the process and escalate the potential issue with OpenReview. \n\nKind Regards,\nMIDL Area Chair for Paper #27\n"}, {"title": "I personally think that performing dMRI-based tissue segmentation is very interesting. But, I am not quite agree with the authors' responses.", "comment": "I personally think that performing dMRI-based tissue segmentation is very interesting. But, I am not quite agree with the authors' responses.\n\n- While there are limited studies (only one as reported), the authors have to propose ways to perform quantitative comparison evaluations; otherwise, the currently reported numbers are not informative. I pointed out that if the existing study used the same data as in the present study: if yes, it at least gives a relatively fair comparison to the numbers reported; if not, the authors could try to implement their method.\n\n- I agree with that using as less data as possible for training is a pros. But, my point here is to include multiple subjects to account for individual variabilities.\n\n- I have no further comments on this point. But, I think some of the replied message should be included in the paper.\n\n- Then, is the segmentation computed by the authors similar to what provided in the HCP database? If not, why the authors think their computed segmentations are better than the preprovided segmentation?"}, {"title": "Answer ", "comment": "Thank you for clarifying the two different steps (point 2) within a spherical CNN (convolution on a sphere in the first layer and grouped convolutions in the following layers).\n\n \n\nRegarding points 1 and 3:\n\nAssuming an FA threshold is utilized to find single fiber voxels, but FAST fails to segment the brain (e.g. infants): How is the label generated for training (apart from manually labeling the FA image)? Here, it would be very important to investigate the effect of transfer learning in more detail.\n\n \n\nA different case would occur if only diffusion data was acquired (or T1/T2 is too noisy): Here, a plausible step would be to use the b0 measurement, since this is, as noted in your response, a low resolution T2 acquisition. Due to this, it would be a really important comparison and should be included in the evaluation to show that your method performs better or at least equivalent.\n\n \n\nRegarding the statement that this is the first approach (or idea) that performs a brain segmentation based on diffusion data, you could also compare your method against approaches presented in https://www.ncbi.nlm.nih.gov/pubmed/23286167 and https://ieeexplore.ieee.org/document/7448418 , since they also perform segmentation of the brain\u2019s microstructure based on machine learning."}], "comment_replyto": ["BJxFNWBc7E", "r1gkazwFQ4", "BJgm2DkJgN", "r1xJMipE7N", "Bygzc23B4V", "Hkxl-F2BEE", "rJlvLKnS4N"], "comment_url": ["https://openreview.net/forum?id=BJgm2DkJgN&noteId=rJl0FdnH4N", "https://openreview.net/forum?id=BJgm2DkJgN&noteId=Hkxl-F2BEE", "https://openreview.net/forum?id=BJgm2DkJgN&noteId=Bygzc23B4V", "https://openreview.net/forum?id=BJgm2DkJgN&noteId=rJlvLKnS4N", "https://openreview.net/forum?id=BJgm2DkJgN&noteId=SkgdWDbUNN", "https://openreview.net/forum?id=BJgm2DkJgN&noteId=S1l3oi7P4V", "https://openreview.net/forum?id=BJgm2DkJgN&noteId=BJe7LTVd4N"], "meta_review_cdate": 1551356564052, "meta_review_tcdate": 1551356564052, "meta_review_tmdate": 1551703116429, "meta_review_ddate ": null, "meta_review_title": "Acceptance Decision", "meta_review_metareview": "Paper 27 - Rejection, due to practical limitations and issues on comparison \n\nThis paper proposes to use spherical CNNs on diffusion data for brain tissue classification. \n\n* R#2 highlights novelty of an so(3) framework but mentions practical limitations, raises issues with comparisons (method and dataset). In a summary, why segmenting on dMRI and not on available structural MRIs. \n\nAuthors indicates challenges in the availability of structural segmentations, and reminds scenarios may arise where structural images may be difficult to segment. \n\n* R#1 finds the method original but has concerns on its evaluation (comparison with other method, used dataset)\n\nAuthors reminds the novelty is to enable learning of diffusion data in a correct SO(3) domain. \n\nR#1 reminds that if an SO(3) framework is proposed, its improvement over a standard Euclidean framework should be evaluated. \n\n* R#3 highlights the methodological novelty on learning diffusion data in SO(3), but has major concerns on methodology (input, architecture) and evaluation (lack of comparison). \n\nAuthors clarify misunderstandings from the reviewer, and refers the reviewer to the original Spherical CNN paper. As a side node, Spherical CNNs were proposed in recent ICML and ICLR papers. The author contribution is on their use on diffusion imaging. \n\nConclusion:\n\nReviewers recommandation are reject-reject-strong reject - mostly due to a practical limitation (why segmenting on diffusion while structure is available) and lack of comparison (what is the improvement over a Euclidean version). All are absolutely certain of their decisions. \n\nGlobal recommendation towards Rejection - but may be considered as a borderline paper depending on paper quotas. Reviewers concern on a missing comparative study is, however, genuine and does require major changes in this submission. \n", "meta_review_readers": ["everyone"], "meta_review_writers": ["MIDL.io/2019/Conference"], "meta_review_reply_count": {"replyCount": 0}, "meta_review_url": ["https://openreview.net/forum?id=BJgm2DkJgN&noteId=H1g25fUHLE"], "decision": "Reject"}