List of Papers By topics Author List
| Paper Info | Reviews | Meta-review | Author Feedback | Post-Rebuttal Meta-reviews |
Authors
Yuqiang Gao, Guanyu Yang, Xiaoming Qi, Yinsu Zhu, Shuo Li
Abstract
Multi-sequence magnetic-resonance-imaging (MRI) images have complementary information that can greatly improve the reliability of diagnosis. However, automated diagnosis of small sample multi-sequence MR images is a challenging task due to: 1) Divergent representation. The difference between sequences and the weak correlation between contained features make the representation extracted from the network tend to diverge, which is profitless to robust classification. 2) Sparse distribution. The small sample size is reflected in the sparse distribution of the prototype, making the network only learn rough demarcation, which is inadequate to medical images with small class interval. In this paper, we propose for the first time a network (SAPJNet) that can adapt to both multi-sequence and small sample conditions, enabling high-quality automatic diagnosis of small-sample multi-sequence MR images, which is of great help to improve clinical diagnostic efficiency. 1) The sequence-adaptive transformer (SAT) of SAPJNet generates joint representations as disease prototypes by filtering intra-sequence features and aggregating inter-sequence features. 2) A prototype optimization strategy (POS) of SAPJNet constrains the prototype distribution by approximating the intra-class prototype and alienating the inter-class prototype. The SAPJNet achieved optimal performance in three tasks: risk assessment of Pulmonary arterial hypertension, classification of idiopathic inflammatory myopathies, and identification of knee abnormalities, with at least a 10%, 10%, and 6.7% improvement in accuracy over all comparison methods.
Link to paper
DOI: https://link.springer.com/chapter/10.1007/978-3-031-16431-6_52
SharedIt: https://rdcu.be/cVD68
Link to the code repository
N/A
Link to the dataset(s)
N/A
Reviews
Review #1
- Please describe the contribution of the paper
this paper proposed to use Transformer and additive-angular-margin loss for small sample multi-sequence MR image classification. However, this paper has several major flaws and fails to illustrate their point:
- The first component, Transformer model, is claimed to filter intra-sequence features and aggregate inter-sequence features, based on attention mechanism. But I didn’t see any details about how to achieve these 2 goals in section 2.1. The overall writing quality is poor and difficult to follow.
- Similarly, neither the section 2.2, prototype optimization strategy, illustrates how to approximate the intra-class prototype and alienate the inter-class prototype. For example, the query sample q^a and support sample s^b corresponding to which modalities, the loss_2 should be explicitly expressed like loss_1, and why you choose the additive-angular-margin loss instead of the ordinary cross-entropy loss for classification.
- Fig 2 is very ambiguous and lacks sufficient explanation.
- Please list the main strengths of the paper; you should write about a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
this paper proposed to use Transformer and additive-angular-margin loss for small sample multi-sequence MR image classification. However, this paper has several major flaws and fails to illustrate their point:
- The first component, Transformer model, is claimed to filter intra-sequence features and aggregate inter-sequence features, based on attention mechanism. But I didn’t see any details about how to achieve these 2 goals in section 2.1. The overall writing quality is poor and difficult to follow.
- Similarly, neither the section 2.2, prototype optimization strategy, illustrates how to approximate the intra-class prototype and alienate the inter-class prototype. For example, the query sample q^a and support sample s^b corresponding to which modalities, the loss_2 should be explicitly expressed like loss_1, and why you choose the additive-angular-margin loss instead of the ordinary cross-entropy loss for classification.
- Fig 2 is very ambiguous and lacks sufficient explanation. E.g., what are the vectors besides the local significance block, how to aggregate into global correlation (through concatenation, pooling, or others), where does the support prototype come from, where are the 2 stages in Prototype optimization strategy, and how to approximate intra-class prototypes and alienate inter-class prototypes.
- The experiment section lacks the explanation of 3 modalities of data (SAX, LAX, and LGE).
- Some minor issues: a. P2, 1st paragraph, the hyperparameter p’s explanation is ambiguous. b. Section 2.2 mentions robust classification. So what kind of attack methods you are dealing with? c. Section 2.2 2nd paragraph says “two outputs of the SAT …” what does the two outputs referring to? d. Still in Section 2.2 2nd paragraph, “The former is reserved for …”. After this, where is the latter one? What’s the purposes of two stages here? e. Table 3 didn’t explain what’s VOT.
- Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.
this paper proposed to use Transformer and additive-angular-margin loss for small sample multi-sequence MR image classification. However, this paper has several major flaws and fails to illustrate their point:
- The first component, Transformer model, is claimed to filter intra-sequence features and aggregate inter-sequence features, based on attention mechanism. But I didn’t see any details about how to achieve these 2 goals in section 2.1. The overall writing quality is poor and difficult to follow.
- Similarly, neither the section 2.2, prototype optimization strategy, illustrates how to approximate the intra-class prototype and alienate the inter-class prototype. For example, the query sample q^a and support sample s^b corresponding to which modalities, the loss_2 should be explicitly expressed like loss_1, and why you choose the additive-angular-margin loss instead of the ordinary cross-entropy loss for classification.
- Fig 2 is very ambiguous and lacks sufficient explanation. E.g., what are the vectors besides the local significance block, how to aggregate into global correlation (through concatenation, pooling, or others), where does the support prototype come from, where are the 2 stages in Prototype optimization strategy, and how to approximate intra-class prototypes and alienate inter-class prototypes.
- The experiment section lacks the explanation of 3 modalities of data (SAX, LAX, and LGE).
- Some minor issues: a. P2, 1st paragraph, the hyperparameter p’s explanation is ambiguous. b. Section 2.2 mentions robust classification. So what kind of attack methods you are dealing with? c. Section 2.2 2nd paragraph says “two outputs of the SAT …” what does the two outputs referring to? d. Still in Section 2.2 2nd paragraph, “The former is reserved for …”. After this, where is the latter one? What’s the purposes of two stages here? e. Table 3 didn’t explain what’s VOT.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Note, that authors have filled out a reproducibility checklist upon submission. Please be aware that authors are not required to meet all criteria on the checklist - for instance, providing code and data is a plus, but not a requirement for acceptance
this paper proposed to use Transformer and additive-angular-margin loss for small sample multi-sequence MR image classification. However, this paper has several major flaws and fails to illustrate their point:
- The first component, Transformer model, is claimed to filter intra-sequence features and aggregate inter-sequence features, based on attention mechanism. But I didn’t see any details about how to achieve these 2 goals in section 2.1. The overall writing quality is poor and difficult to follow.
- Similarly, neither the section 2.2, prototype optimization strategy, illustrates how to approximate the intra-class prototype and alienate the inter-class prototype. For example, the query sample q^a and support sample s^b corresponding to which modalities, the loss_2 should be explicitly expressed like loss_1, and why you choose the additive-angular-margin loss instead of the ordinary cross-entropy loss for classification.
- Fig 2 is very ambiguous and lacks sufficient explanation. E.g., what are the vectors besides the local significance block, how to aggregate into global correlation (through concatenation, pooling, or others), where does the support prototype come from, where are the 2 stages in Prototype optimization strategy, and how to approximate intra-class prototypes and alienate inter-class prototypes.
- The experiment section lacks the explanation of 3 modalities of data (SAX, LAX, and LGE).
- Some minor issues: a. P2, 1st paragraph, the hyperparameter p’s explanation is ambiguous. b. Section 2.2 mentions robust classification. So what kind of attack methods you are dealing with? c. Section 2.2 2nd paragraph says “two outputs of the SAT …” what does the two outputs referring to? d. Still in Section 2.2 2nd paragraph, “The former is reserved for …”. After this, where is the latter one? What’s the purposes of two stages here? e. Table 3 didn’t explain what’s VOT.
- Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review: https://conferences.miccai.org/2022/en/REVIEWER-GUIDELINES.html
this paper proposed to use Transformer and additive-angular-margin loss for small sample multi-sequence MR image classification. However, this paper has several major flaws and fails to illustrate their point:
- The first component, Transformer model, is claimed to filter intra-sequence features and aggregate inter-sequence features, based on attention mechanism. But I didn’t see any details about how to achieve these 2 goals in section 2.1. The overall writing quality is poor and difficult to follow.
- Similarly, neither the section 2.2, prototype optimization strategy, illustrates how to approximate the intra-class prototype and alienate the inter-class prototype. For example, the query sample q^a and support sample s^b corresponding to which modalities, the loss_2 should be explicitly expressed like loss_1, and why you choose the additive-angular-margin loss instead of the ordinary cross-entropy loss for classification.
- Fig 2 is very ambiguous and lacks sufficient explanation. E.g., what are the vectors besides the local significance block, how to aggregate into global correlation (through concatenation, pooling, or others), where does the support prototype come from, where are the 2 stages in Prototype optimization strategy, and how to approximate intra-class prototypes and alienate inter-class prototypes.
- The experiment section lacks the explanation of 3 modalities of data (SAX, LAX, and LGE).
- Some minor issues: a. P2, 1st paragraph, the hyperparameter p’s explanation is ambiguous. b. Section 2.2 mentions robust classification. So what kind of attack methods you are dealing with? c. Section 2.2 2nd paragraph says “two outputs of the SAT …” what does the two outputs referring to? d. Still in Section 2.2 2nd paragraph, “The former is reserved for …”. After this, where is the latter one? What’s the purposes of two stages here? e. Table 3 didn’t explain what’s VOT.
- Rate the paper on a scale of 1-8, 8 being the strongest (8-5: accept; 4-1: reject). Spreading the score helps create a distribution for decision-making
4
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
this paper proposed to use Transformer and additive-angular-margin loss for small sample multi-sequence MR image classification. However, this paper has several major flaws and fails to illustrate their point:
- The first component, Transformer model, is claimed to filter intra-sequence features and aggregate inter-sequence features, based on attention mechanism. But I didn’t see any details about how to achieve these 2 goals in section 2.1. The overall writing quality is poor and difficult to follow.
- Similarly, neither the section 2.2, prototype optimization strategy, illustrates how to approximate the intra-class prototype and alienate the inter-class prototype. For example, the query sample q^a and support sample s^b corresponding to which modalities, the loss_2 should be explicitly expressed like loss_1, and why you choose the additive-angular-margin loss instead of the ordinary cross-entropy loss for classification.
- Fig 2 is very ambiguous and lacks sufficient explanation.
- Number of papers in your stack
4
- What is the ranking of this paper in your review stack?
4
- Reviewer confidence
Very confident
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
Not Answered
- [Post rebuttal] Please justify your decision
Not Answered
Review #2
- Please describe the contribution of the paper
This paper proposes a new deep learning method for optimizing disease diagnoses that rely on small sample multi-sequence magnetic resonance imaging (MRI) data. Specifically, using a common neural network known as ResNet50 as the backbone, the study has developed a SAPJNet approach that: 1) behaves like a sequence-adaptive transformer to generate joint feature representations of disease prototypes, and 2) constrains the prototype distribution through a prototype optimization strategy. Experiments using MRI of heart and knee show that the SAPJNet is 6-10% more accurate than four other related neural networks.
- Please list the main strengths of the paper; you should write about a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1) Developed a new neural network approach to deal with common clinical MRI issues 2) Compared the proposed method with several other methods in the literature 3) Added a brief ‘ablation study’ to gain insight of the proposed technique
- Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.
1) Lack of details in Method 2) The output of the networks is unclear 3) Lack of justification of the steps chosen in image preprocessing, and in some parameters chosen in Method 4) A few grammar issues
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Note, that authors have filled out a reproducibility checklist upon submission. Please be aware that authors are not required to meet all criteria on the checklist - for instance, providing code and data is a plus, but not a requirement for acceptance
No mentioning about it, except listing of sources for the comparison methods
- Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review: https://conferences.miccai.org/2022/en/REVIEWER-GUIDELINES.html
This is an interesting study addressing important questions associated with deep learning using real-world medical imaging data, such as MRI. The proposed method appears to be able to overcome 2 critical limitations of clinical MRI: small sample size, and multiple sequence acquisitions. However, there are a few questions about the paper that deserves further attention as seen below.
1) The method is largely unclear. For example, in section 2.1, how ‘the SAT accurately extracts features and generates prototypes that mimic a doctor’s overall assessment of …’? How are ‘these representations aggregated and translated into new semantics’? Further, how are the parameters ‘p’ and ‘delta’ set and what do they mean?
2) In section 2.2, how are the positive and negative sample pairs calculated; what does the ‘supervisory information’ refer to; and what is the relationship between ‘pre-prototype’ and ‘prototype’? In addition, while there is an equation (#2), adding additional explanation for the ‘additive-angular-margin loss’ would help.
3) In Fig. 2, how was the ‘filter intra-sequence features’ step done? What’s the usage of ‘padding’ here – expand the cropped images to 90x90? How to ‘approximate intra-class prototypes’ and ‘alienate inter-class prototypes?
4) In Experiments and Results (section 3), the output of the proposed and comparison networks are not defined; multiple abbreviations, ‘PAH’, ‘IIM’, ‘LGE’ et al, are not defined at first use.
5) Also in the above section, it says that ‘different sequences of the same patient, although …, were not spatially aligned’. Is this beneficial or harmful, and why? Related to this point, in image preprocessing, the MR images do not seem to be normalized, in any dataset. How would that impact network performance, and why is that preferred?
6) In Fig. 3, it is unclear what ‘abnormalities’ are supposed to be seen despite the use of arrows in the panels. Including an example of normal image would also help.
7) In section 3.3 (page 8), it is unclear what this part means: ‘As shown in Fig. 4, compared with the baseline method, it can be seen that with the reduction of training samples, the performance of the SAPJNet is better, and its performance loss is smaller in the five training’. In Fig. 4 (bottom plots), it appears that the sample size is increasing from 20 to 40 instead of ‘decreasing’ as the network performance increases. Please verify.
- Rate the paper on a scale of 1-8, 8 being the strongest (8-5: accept; 4-1: reject). Spreading the score helps create a distribution for decision-making
5
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This paper addresses 2 common and important questions facing the use of clinical MRI in deep learning. The proposed method shows reasonable result compared to several other approaches. However, the study has several limitations as mentioned above.
- Number of papers in your stack
4
- What is the ranking of this paper in your review stack?
3
- Reviewer confidence
Confident but not absolutely certain
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
Not Answered
- [Post rebuttal] Please justify your decision
Not Answered
Review #3
- Please describe the contribution of the paper
The authors introduce a sequence adaptive transformer based architecture that generates joint representations according to disease prototype. The paper deals with how to generalize disease classification from MR images despite the small sample size and presence of multiple sequences.
- Please list the main strengths of the paper; you should write about a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
Different sequences of MR images depicting structure, motion are utilized effectively for the classification purpose. Redundant intra-sequence features have been filtered using self-attention while inter-sequence feature correlations have been explored. Instead of simple fusion, multiple sequence features have been fused effectively. The authors introduce a contrastive learning inspired loss function to better classification performance even in sparse data scenarios. The paper groups like and unlike pairs according to disease category and alienates similar class prototypes. The results section is nicely presented.
- Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.
1) Why did the authors choose cosine similarity as a correlation method? Also it would be great to have some ablation in terms of losses employed in alienating features. Please compare against other naive similarity losses - like modified Barlow twins, SimSiam etc
2) The authors came up with a transformer-based learning paradigm. Can you please explain 2.1 in more details - especially how the self-attention is modeled to filter intra-sequence features. A few mathematical equations would be great. The local significance section is vague.
3) Since the authors called it a Sequence adaptive transformer, what will happen if one or two sequences are missing for few patients but available for others? Like motion is available for one and not for another. It’s a common and practical scenario for MRI input.
4) I get how the authors ‘approximate intra-class prototypes’ and ‘alienate inter-class prototypes’ through the loss function. But it would be good if they can give a more clear explanation regarding it. Also, only their mention without any context in Fig 2 makes it a bit hard to understand.
- Please rate the clarity and organization of this paper
Very Good
- Please comment on the reproducibility of the paper. Note, that authors have filled out a reproducibility checklist upon submission. Please be aware that authors are not required to meet all criteria on the checklist - for instance, providing code and data is a plus, but not a requirement for acceptance
The authors have provided necessary parameter deatils. They intend to make the code public.
- Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review: https://conferences.miccai.org/2022/en/REVIEWER-GUIDELINES.html
Please address the issues listed in weakness.
- Rate the paper on a scale of 1-8, 8 being the strongest (8-5: accept; 4-1: reject). Spreading the score helps create a distribution for decision-making
6
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper is well written. The methodology proposed to handle sparse distribution and mylti-sequence images is novel. They utilize SOTA attention mechanism and contrastive learning in different stages of their architecture. The results show a justifiable improvement when compared against other methods.
- Number of papers in your stack
5
- What is the ranking of this paper in your review stack?
1
- Reviewer confidence
Confident but not absolutely certain
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
Not Answered
- [Post rebuttal] Please justify your decision
Not Answered
Primary Meta-Review
- Please provide your assessment of this work, taking into account all reviews. Summarize the key strengths and weaknesses of the paper and justify your recommendation. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
This paper proposed to use Transformer and additive-angular-margin loss for small sample multi-sequence MR image classification. This paper received two positive reviews and one negative reviews. In the rebuttal, the author should clarify the concerns raised by review 1, including the detailed description of prototype optimization strategy, more illustration of Fig 2 and the first component.
- What is the ranking of this paper in your stack? Use a number between 1 (best paper in your stack) and n (worst paper in your stack of n papers). If this paper is among the bottom 30% of your stack, feel free to use NR (not ranked).
6
Author Feedback
Thanks to the meta reviewer and all reviewers for their meticulous review, we appreciate your recognition of the study on: 1) Novel innovation: “Overcome 2 critical limitations of clinical MRI.” -R2. “The methodology is novel.” -R3. 2) Effective design: “Different sequences are utilized effectively.” -R3. “Multiple sequence features have been fused effectively.” -R3. 3) justifiable improvement: “Give a brief ablation study to gain insight of the proposed technique.” -R2. “The results nicely show a justifiable improvement.” -R3.
Q1: About Network: “How to filter intra-sequence features and aggregate inter-sequence features?” -R1R2R3. “What are two outputs of SAT?” -R1R2. A1: -The network 1) filters intra-sequence features by inhibiting ineffective features with the relevance score to others, and 2) aggregates inter-sequence features by linearly summing all weighted features into one learnable token. The two steps fully fuse effective features while keeping the prototype low-dimensional. They are typical of Transformer and are elaborated by many basic studies such as ViT. Therefore, the description in Sec. 2.1 focuses on the adaption of Transformer to multi-sequences rather than the principle. -Outputs of SAT are: 1) the pre-prototype that directly splices ResNets’ output features, 2) the prototype that is the learnable token.
Q2: About Optimization: “How to approximate intra-class prototypes and alienate inter-class prototypes in two stages?” -R1R2. “How are positive and negative pairs calculated and which modality do q^a, s^b correspond to?” -R1R2. “Why choose Metric learning and Additive-angular-margin Loss over Cross-entropy Loss?” -R1R3. A2: -Two types of metric learning are used to approximate or alienate prototypes. In Stage1, the pair-based Cosine-similarity Loss increases intra-class prototypes’ similarity and decreases inter-class prototypes’ similarity. In Stage2, the proxy-based Additive-angular-margin Loss learns the cluster center of prototypes. -Pairs are constructed using sparse sampling methods, where a and b represent different shots rather than modalities. It applies to all modalities and is detailed in para. 2 of Sec. 2.2. -Because metric learning solves the small-sample problem by extending the training set in disguise. Additive-angular-margin Loss maintains consistency in measuring prototypes rather than mapping them to the label space, which is shown to be better in the ablation study.
Q3: About Fig. 2: “What are vectors besides the local significance block, where does the support prototype come from and where are the 2 stages in POS?” -R1, “What’s the ‘padding’ here?” -R2. A3: -The vectors besides the local significance block are ResNets’ output features. Supporting prototypes are the product of support samples encoded by SAT. Two stages of POS are demonstrated by characterizing two typical metric learning styles. All three questions are easily understood in conjunction with Fig. 2 and Sec. 2. -The “padding” is only in the appendix but not in Fig. 2. It represents making data block consistent in the slice dimension.
Q4: About Robustness: “What attack you are dealing with?” -R1. A4: Robust classification mentioned in Sec. 2.2 means the tolerance for reduced training samples, as shown in Fig. 4.
Q5: About Fig. 4: “The performance seems better as the sample size increases not decreases?” -R2. A5: Fig.4 aims to prove that our method’s performance degrades less than others as the sample size decreases and is therefore better.
Q6: About Hyperparameters: “What do the parameters p and delta mean and how are they set?” -R2. A6: The only hyperparameter p is set to 4. The meaning and setting, and relationship of parameters are detailed in para. 2 of Sec. 2.1.
Q7: About Preprocessing: “Why is the spatial misalignment of sequences and the seeming lack of normalization?” -R2. A7: Because ResNets’ output features are aligned at the feature level, and all images are implicitly normalized by BatcheNorm on input.
Post-rebuttal Meta-Reviews
Meta-review # 1 (Primary)
- Please provide your assessment of the paper taking all information into account, including rebuttal. Highlight the key strengths and weaknesses of the paper, clarify how you reconciled contrasting review comments and scores, indicate if concerns were successfully addressed in the rebuttal, and provide a clear justification of your decision. If you disagree with some of the (meta)reviewer statements, you can indicate so in your meta-review. Please make sure that the authors, program chairs, and the public can understand the reason for your decision.
In the rebuttal, the author addressed the comments of reviewers well, including the prototype issues, and illustrations of fig,2. Reviewers also admit the novelty of this paper. I would suggest to accept this paper after rebuttal.
- After you have reviewed the rebuttal, please provide your final rating based on all reviews and the authors’ rebuttal.
Accept
- What is the rank of this paper among all your rebuttal papers? Use a number between 1/n (best paper in your stack) and n/n (worst paper in your stack of n papers). If this paper is among the bottom 30% of your stack, feel free to use NR (not ranked).
4
Meta-review #2
- Please provide your assessment of the paper taking all information into account, including rebuttal. Highlight the key strengths and weaknesses of the paper, clarify how you reconciled contrasting review comments and scores, indicate if concerns were successfully addressed in the rebuttal, and provide a clear justification of your decision. If you disagree with some of the (meta)reviewer statements, you can indicate so in your meta-review. Please make sure that the authors, program chairs, and the public can understand the reason for your decision.
This paper proposed a Transformer based model to fuse multi-sequence MRI input and a prototype optimization strategy dealing with the sparse input via metric learning. In my mind, this is a borderline paper. It has moderate novelty and shows significant performance improvement from the competitive methods. On the other hand, claiming self-attention to be able to filter intra-sequence features may be over-optimistic, and lacks insights from the experiment to support. Also, the original paper is not very easy to follow with many technical details missing. If this paper is eventually accepted, I strongly suggest to incorporate the clarification about the details of metric learning into the final paper.
- After you have reviewed the rebuttal, please provide your final rating based on all reviews and the authors’ rebuttal.
Accept
- What is the rank of this paper among all your rebuttal papers? Use a number between 1/n (best paper in your stack) and n/n (worst paper in your stack of n papers). If this paper is among the bottom 30% of your stack, feel free to use NR (not ranked).
9
Meta-review #3
- Please provide your assessment of the paper taking all information into account, including rebuttal. Highlight the key strengths and weaknesses of the paper, clarify how you reconciled contrasting review comments and scores, indicate if concerns were successfully addressed in the rebuttal, and provide a clear justification of your decision. If you disagree with some of the (meta)reviewer statements, you can indicate so in your meta-review. Please make sure that the authors, program chairs, and the public can understand the reason for your decision.
The paper proposed an approach to fuse the modes of MRI used during imaging (T1, T2, Flair, DTI, etc.) in order to improve the diagnosis. This is an important problem for which the authors have presented an elegant solution. The reviewers are positively inclined as are other ACs after reading the rebuttal.
- After you have reviewed the rebuttal, please provide your final rating based on all reviews and the authors’ rebuttal.
Accept
- What is the rank of this paper among all your rebuttal papers? Use a number between 1/n (best paper in your stack) and n/n (worst paper in your stack of n papers). If this paper is among the bottom 30% of your stack, feel free to use NR (not ranked).
7