List of Papers By topics Author List
| Paper Info | Reviews | Meta-review | Author Feedback | Post-Rebuttal Meta-reviews |
Authors
Minyan Zeng, Yutong Xie, Minh-Son To, Lauren Oakden-Rayner, Luke Whitbread, Stephen Bacchi, Alix Bird, Luke Smith, Rebecca Scroop, Timothy Kleinig, Jim Jannes, Lyle J Palmer, Mark Jenkinson
Abstract
While acute ischemic stroke due to large vessel occlusion (LVO) may be life-threatening or permanently disabling, timely intervention with endovascular thrombectomy (EVT) can prove life-saving for affected patients. Appropriate patient selection based on prognostic prediction is vital for this costly and invasive procedure, as not all patients will benefit from EVT. Accurate prognostic prediction for LVO presents a significant challenge. Computed Tomography Perfusion (CTP) maps can provide additional information for clinicians to make decisions.
However, CTP maps are not always available due to variations in available equipment, funding, expertise and image quality. To address these gaps, we test (i) the utility of acquired CTP maps in a deep learning prediction model, (ii) the ability to improve flexibility of this model through image synthesis, and (iii) the added benefits of including multi-task learning with a simple clinical task to focus the synthesis on key clinical features. Our results demonstrate that network architectures utilising a full set of images can still be flexibly deployed if CTP maps are unavailable as their benefits can be effectively synthesized from more widely available images (NCCT and CTA). Additionally, such synthesized images may help with interpretability and building a clinically trusted model.
Link to paper
DOI: https://doi.org/10.1007/978-3-031-43904-9_67
SharedIt: https://rdcu.be/dnwId
Link to the code repository
N/A
Link to the dataset(s)
N/A
Reviews
Review #3
- Please describe the contribution of the paper
In this paper, the authors describe a pair of model architectures useful for image synthesis and functional outcome prediction, respectively, in patients with AIS. In the former case, NCCT and CTA imaging is used to synthesize perfusion parameter maps. In the latter case, spatial data (images or maps) is processed to extract a single scalar feature which is then used in tandem with clinical variables to predict binarized mRS via logistic regression. The technical novelty of the proposed method lies in the explicit synthesis of CTP maps from NCCT and CTA using a hybrid loss function that also incorporates a clinical task. Based on the authors’ results, which show that both proposed models outperform comparison methods for mRS prediction, they conclude that their proposed workflow helps to extract clinically-relevant features from NCCT+CTA.
- 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.
Synthesizing maps of temporally-sensitive parameters from a single-timepoint image is a novel idea that, at face value, would seem to be an impossible task. However, based on the authors’ description, I believe that their method does not synthesize the true parameter maps (with biologically interpretable values) but rather a normalized version of each map with scalar values constrained to a fixed range. From this perspective, the image synthesis pathway could be viewed as an image translation model that also acts as a feature extractor for functional stroke outcome prediction. This is particularly interesting because it makes the overall model a rare example of hierarchical multi-task learning in AIS outcome prediction. As models developed to unify aspects of biological (tissue-level) and functional modelling in AIS become increasingly popular, these types of multi-task learning studies will only become more relevant. Furthermore, the aspect of image-to-image translation present in this study may have implications for the imputation of missing images in multi-modal stroke image analysis, but this is not entirely clear based on the current text of the manuscript.
The authors have also taken care to evaluate their models using multiple comparison methods at both stages of the multi-task hierarchy. This allows the separate analysis of their feature extraction (image synthesis) network and their overall functional outcome prediction pipeline which greatly benefits the quality and interpretability of their results.
- 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.
Some aspects of the justification for the study seem to be at odds with its methods. In particular, the authors hypothesize that “models incorporating CTP maps can be improved through image synthesis from commonly available image modalities (e.g., NCCT and CTA).” However, it does not appear that a model which incorporates true CTP maps was evaluated on synthesized data in this study. Rather, the prognostic prediction model was trained and evaluated solely on synthesized data which, though still interesting and relevant to stroke outcome prediction, does not necessarily require the synthesized images to closely resemble true CTP maps in order to have prognostic value. The authors also seem to suggest that their image synthesis model is distinct from a general image-to-image translation model by virtue of including a joint task in the optimization process. However, it is not uncommon for medical image-to-image translation models to include joint tasks such as segmentation (doi: 10.1109/TMI.2020.2974159) or the preservation of lesion-related imaging features (doi: 10.1007/s11548-022-02828-4). Therefore, the use of multi-task learning is not especially novel in the domain of image synthesis; however, the combination of multi-task learning and image synthesis is novel in the domain of stroke modeling and this could perhaps be emphasized.
Overall, several aspects of the data preprocessing and experimental analysis are not sufficiently described, including the perfusion analysis (and deconvolution, if used), training targets and ground truths, and model training and inference. Specific examples are provided in box 9: constructive feedback, but do not fit in the current textbox due to character limit restrictions.
Finally, some of the authors’ conclusions do not necessarily appear to be supported by the experimental results. For instance, the authors suggest a significant “overlap between the features that can be learned directly from NCCT and CTA and those from non-imaging data.” However, the model that used only the non-imaging data (AUC ROC = .773) vastly outperformed the model which used only direct NCCT+CTA (AUC ROC = .571), which was very close to chance level for the prediction of binary mRS (AUC ROC = 0.5). My interpretation of these results would rather be that very few useful prognostic features are represented directly by the NCCT+CTA and that the non-imaging data contains many valuable prognostic features not directly present in the NCCT+CTA. Additionally, the authors conclude that the proposed framework can “aid in the selection of patients for high-stakes time-critical EVT” and “improve confidence in building a clinically trusted model.” However, AI-guided patient selection for EVT and related clinician attitudes were not measured in this study and should not be phrased as an interpretation of the study’s results, although they could certainly be provided as avenues to explore in future studies.
- 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
I would not be able to reproduce this study due to notable omissions in the training and inference process for the comparison methods that this study evaluated.
- 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/2023/en/REVIEWER-GUIDELINES.html
To align the experiment with its justification, I can think of two possible recommendations. If the authors’ intended use case for their model is to enable techniques designed for CTP to function for patients having only NCCT+CTA, then I would expect to see the synthesized CTP used for inference in a model that was trained on the true CTP maps. Alternatively, if the authors’ intended use case is for their image synthesis pathway to act as a feature extractor for functional outcome prediction (and therefore produce the best possible functional outcome predictions in a model that was optimized to receive NCCT+CTA instead of CTP for inference), then I would expect to see a comparison between the proposed method and an end-to-end model with roughly the same number of parameters for predicting binarized mRS. The reason for this is that the results in figures 3 and 4 seem to come from a workflow that does not include CTP synthesis and therefore uses only the relatively shallow ‘stage two’ model. In this case, it is not especially fair to compare raw NCCT+CTA to synthesized CTP maps when only the latter had the benefit of passing through the generator for convolutional feature extraction. Phrased another way, it is exciting that prognostic features can be extracted from NCCT+CTA by optimizing a network for image-to-image translation, but are these features better than what could be extracted by an equivalent network optimized directly for functional outcome prediction?
More to this point, I think that this paper lacks a justification for why CTP maps are assumed to be the best prognostic features for functional outcome prediction (and therefore the target of the image synthesis pathway). As the authors state, it is true that CTP maps are used clinically to identify the core and penumbra which have some biological relevance, but in this case, why not attempt to predict the infarct and penumbra directly from the NCCT+CTA as in (doi: 10.1177/0271678X211023660)? There is also the school of thought that core and penumbra are rather an oversimplification of stroke etiology (doi: 10.1161/STROKEAHA.120.030620) and that deep convolutional neural networks should be allowed to optimize their own features, which may differ from those used by human clinicians, for the task at hand (doi: 10.1016/j.media.2022.102610). This is not to say that enforcing the use of synthesized CTP maps for functional outcome prediction is necessarily a bad idea, but I do think that the paper requires either a theoretical or an experimental justification for why this approach was chosen instead of the potentially more-viable alternatives described above.
Regarding the methods section, some key details regarding the data are missing or ambiguous. In particular, I would be interested to read about the perfusion analysis, including the selection of the AIF and deconvolution algorithm if applicable. If CTP maps were approximated directly from tissue curves rather than the deconvolved residual curves, a justification would be warranted. The process for training and evaluating the models is also not sufficiently described. For example, the authors refer to the left brain / right brain classification task used in the loss function of their ‘stage 1’ model as the ‘simple clinical task’ or ‘clinically-relevant task,’ but do not explain what this task actually entails. Because stroke hemisphere is listed in section 2.1 as a clinical feature, it is my assumption that the ‘clinical’ task was to identify the hemisphere ipsilateral to the occlusion, but this should be explicitly stated.
Additionally, the authors describe their overall proposed model in great detail in section 2.2., but do not describe how the additional experiments related to input image modality, image synthesis model, and other published mRS prediction methods were conducted. In particular, the sentence “images were input into the models with the architecture described in section 2.2.” is not useful to the reader, since section 2.2 describes two different models with image-based inputs at multiple locations and no description of how image modalities other than NCCT+CTP might be incorporated. This has serious implications for the interpretation of the results. For example, the stage two model (as described) was trained only using the synthetic perfusion maps produced by the stage one model. Was this same trained instance of the stage two model also used for the evaluation of each alternative synthesis model (UNET, WGAN, CycleGAN, L2GAN) and combination of input image modalities (NCCT+CTA, CTP, NCCT+CTA+CTP), or were eight separate instances of model two trained from scratch using their respective inputs? In the former case, the training target for the alternative synthesis models (the true CTP maps) is different from the input used to optimize the stage two model (the output of the stage one model), which would negatively impact performance. In the latter case, different models are used to predict mRS from the true and synthesized CTP, so it becomes difficult to support the authors’ claim that “the proposed method can recover most of its predictive ability when CTP maps are unavailable.” This claim would seem to imply that the synthesized CTP maps could be used in a model for which true CTP maps are typically available and used. Rather, the experimental results would show that a model which has never seen true CTP maps can perform reasonably well on surrogates derived from NCCT+CTA regardless of how closely they resemble true CTP maps.
Generally in this section, I would advise that too much of the paper is spent simply listing the architecture of the model. If this model closely resembles one that is previously published (L2GAN), then I would consider it sufficient to cite this model and refer the reader to an appendix or supplementary materials for specific details on the kernel size and layer sequence. It would be much more valuable to the reader, in my opinion, to have this space dedicated to describing the experimental procedure.
Regarding the interpretation of the authors’ results, I would recommend to include some statistical analyses to indicate the extent to which key experimental differences may be the result of random error rather than experimental manipulations. For example, it is possible to compare ROC curves using many statistical analysis softwares, or even potentially by hand (https://www.ncss.com/wp-content/themes/ncss/pdf/Procedures/NCSS/Comparing_Two_ROC_Curves-Paired_Design.pdf). I also find it dubious that a simple logistic regression consisting of only clinical variables processed by a single network node (AUC ROC = .773) outperformed all of the state-of-the-art deep learning models which themselves use the same clinical variables as well as imaging (AUC ROC = .746 to .766). In addition to analyzing these experimental results for statistical significance, it may be beneficial to verify these methods’ implementation. AI-guided patient selection for EVT and related clinician attitudes were not measured in this study and should not be phrased as an interpretation of the study’s results. Finally, in reference to the images generated by the stage one model, I would suggest to note that the synthesized maps do not contain the range of scalars typical of semi-quantitative CTP maps and therefore would not necessarily perform well in all models or contexts that expect raw CTP maps as input.
Other minor feedback: – Intro line 3: ‘irretrievable’ is an odd word choice – References are easier to parse when they are numbered sequentially in order of in-text appearance – Page 2 line 7: ‘hypodensity’ should read ‘hypodense’ – 2.1. Dates are ambiguous as to whether they are formatted as dd/mm/yyyy or mm/dd/yyyy – Section 2.2 lists clinical variables not present in section 2.1. Please list all clinical variables in section 2.1
- 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 shows novelty in its approach to functional outcome prediction in CTP. However, the clinical relevance of this novelty is not well-justified in the current text. The paper’s technical approach uses familiar model structures and the well-established principles of multi-task and joint-task learning, so a strong theoretical or experimental justification for the specific application of these principles to functional outcome prediction in AIS is rather important. Additionally, it is difficult to comment on the quality and rigor of the experiment as well as the soundness of the authors’ analysis of their results due to notable omissions in the methods section of the paper. I see the potential for a very interesting paper, especially considering the large number of patient datasets used and the considerable effort required to evaluate so many comparison methods. However, considerable changes to the manuscript are required to communicate the value of this experiment clearly.
- Reviewer confidence
Very confident
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
N/A
- [Post rebuttal] Please justify your decision
N/A
Review #4
- Please describe the contribution of the paper
The authors develop generative methods for generating CTP data (4 perfusion maps) per patient using NCCT and CTA data and some patient, clinical data. A prediction model is also trained and tested to determine if or how the synthetic CTP data can be prognostic of longer-term patient outcomes, using dichotomized mRS as the metric.
- 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.
*Any additional, trusted, information that can be used by a stroke team in the early hours of an ischemic is valuable. As noted by the authors, there are access and methodological challenges with CT Perfusion as a tool across a heterogeneous stroke network. So, the ability to generate artifacts with additional diagnostic information, and in a familiar form like CTP maps is very powerful. *The usage of both imaging and clinical data in the synthesis and prognostic stages is very important, and is not widely done in other neuroradiology DL work, thus this is an important contribution to the literature.
- Affine registration of the CT data to a template - important pre-processing. *
- 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.
*methodology; unclear what outputs are created in the first-step, the non-imaging variables used and how they are presented to the logistic regression. *The primary weakness of the work is the additional uncertainty in the prognostic value because synthetic data, with their own uncertainty, are injected into the pipeline. It will be difficult, I think, to tract the effect on prognostic ‘goodness’ because errors in the synthetic CTP may/will have unknown influences in the 2nd stage. There is a good attempt to address that issue by showing the performance using acquired/actual CTP maps as inputs (tables 2 and 3). CTP, itself, can be unreliable and ‘noisy” as the authors mention. So, it may be hard to understand how errors in the synthetic CTP manifest for what type of patients. I believe a lot of clinical validation and testing would be required to demonstrate the robustness of the prognostic output. But, it is compelling work.
- 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
Even with access to the dataset, it would be moderately difficult to reproduce. The biggest challenge is the 2nd stage - how are the outputs of the first model in stage2 combined with which clinical data.
- 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/2023/en/REVIEWER-GUIDELINES.html
Pg 4 - “To perform the above tasks, we propose a two-stage…” is confusing wording. I am not sure if you mean a ‘global’ 2 stage framework or that there are 2 stages for each sub-task (map creation and prognostication). Pg5 - clinical-guided synthesis: The description of the discriminator’s design doesn’t match well Figure 2. It’s unclear in Figure 2 where/how the FC layers of discrimination occur and that the ‘real’ CTP maps are fed to it. The synthetic CTP maps aren’t considered in the Loss function? Pg6 - assuming you used the same data for all the tests? did you implement the models of 2,14? Section 3.1 - hard to determine the overlap potential between the NCCT, CTA and clinical features without knowing the clinical data types.
- 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 perfection of synthetic CTP maps (given robust testing and acceptance) can be a very powerful tool in the management of stroke - especially in under-resources areas.
- Reviewer confidence
Very confident
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
N/A
- [Post rebuttal] Please justify your decision
N/A
Review #5
- Please describe the contribution of the paper
This paper proposes a method of data synthesis of computed tomography perfusion maps for the prognostic outcome of stroke patients. These are generated from a noisier and less time consuming imaging modality (Non-Contrast Computed Tomography and CT angiography). Through a series of ablation studies they demonstrate the usefulness of CTP maps for predicting patient outcomes compared to non-imaging and NCCT+CTA imaging modalities and show that CTP maps are indeed useful for outcome prediction. They then use synthesized CTP maps as an additional modality to NCCT+CTA and show improved prognostic predictions over NCCT+CTA alone.
- 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.
Methodology: The author’s justify the use of this method through ablation and comparison to several baseline methods. The motivation for synthesizing the CTP maps is clear and missing data is relevant in many clinical contexts. The methodology is fairly simple and the augmentation of the GAN with an additional classification loss does seem to improve results. The tables are extensive and cover other state of the art methods. Clarity: Overall the paper is well written, with clear descriptions of the data, training process, motivation and results.
- 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.
Novelty: Missing modality synthesis using a UNET GAN is limited in its novelty( e.g. Multimodal MRI Synthesis Using Unified Generative Adversarial Networks).
- Please rate the clarity and organization of this paper
Excellent
- 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 code would be available and the procedure is well described. Additional implementation information was available in the Appendix.
- 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/2023/en/REVIEWER-GUIDELINES.html
No comments.
- 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?
A strong paper on missing modality synthesis which is limited by its lack of novelty. The justification and results are clear and impressive but doesn’t contribute new methodology.
- 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
6
- [Post rebuttal] Please justify your decision
The rebuttal has not changed my review. The paper is well motivated as it synthesizes a difficult to obtain modality and provides a slight increase in the prognostic capabilities of a network when using the synthetic image.
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.
The authors propose a deep learning model that can effectively utilize both acquired CTP maps and synthesized images when CTP maps are unavailable. They also demonstrate the added benefits of including a clinical task in the model training to focus on key clinical features. The results show that synthesized images can improve the flexibility and generalization of the model. The paper is very interesting, however reviewers raised some important concerns and suggestions that need to be addressed. The major concerns are:
- Is the method synthesizing true CTP maps or a surrogate parameter map for AIS outcome prediction.
- Why non-imaging data outperform NCCT+CTA imaging features and what is the clinical benefits of using imaging features in this case?
- A theoretical or an experimental justification for why this prediction CTP maps from NCCT+CTA is more clinically useful compared to using deep learning to predict outcome directly from NCCT+CTA?
Author Feedback
We thank all reviewers and the meta-reviewer for their helpful and thorough feedback. We appreciate that all reviewers acknowledge the importance of our work (e.g., R3: “I see the potential for a very interesting paper”; R4: “This is an important contribution to the literature.”). In the following, we address comments grouped by theme:
- Main Objective (AC, R3) We believe our main objective was not clearly described. It was to test whether CTP maps, as used clinically in our hospital, were also beneficial for outcome prediction with deep learning. Following initial success, we tested whether similar information could be extracted from NCCT and CTA, which are more widely available. We chose to explore synthesizing CTP maps, which we believe acts as a type of auxiliary task, guiding the network to extract more clinically-relevant features, especially when data is limited. This is our theoretical justification for our approach and the observed results. It also has potential benefits for missing data and explainable AI, to help enhance trust, though these were not explicitly tested, and we will modify our text to clarify this.
- CTP Maps (AC, R3) The CTP maps synthesized by our method were linearly scaled versions of the “true” (quantitative) maps. This reflected the fact that our clinical collaborators made decisions using the within-scan relative values, rather than absolute values that may be influenced by a range of nuisance factors (e.g. head size, blood pressure). We used CTP maps generated by the scanner (Canon), as used clinically. The reviewer correctly states that predicting core infarct and penumbra is a reasonable alternative, but it has often been explored. Instead, we chose to explore synthesis of CTP maps, thus avoiding arbitrary thresholds or manual labelling, and leveraging the full information content of the CTP maps.
- Results and Interpretation (AC, R3, R4) Clinical (non-imaging) data alone did outperform basic NCCT+CTA, but it did worse than NCCT+CTA+CTP (either synthetic or acquired CTP). This demonstrates that imaging is beneficial and results using NCCT+CTA acquisitions, augmented by synthetic CTP, were close to the results using acquired CTP. We believe we were unclear about the overlap of information between clinical (non-imaging) and NCCT+CTA data. Our intention was to point out that the basic network with NCCT+CTA did not extract features that usefully added to the clinical data, although this was not reciprocal since clinical data did contain useful information beyond what this basic network extracted from NCCT+CTA. However, NCCT+CTA can supplement the clinical data as the synthetic CTP maps are derived from NCCT+CTA and hence information is in there, so we will remove the statement about the overlap as we agree that it is unhelpful/confusing. Clinical (non-imaging) data alone did outperform other SOTA models but this may be due to how each model chose to combine these with the imaging information, such that in some cases useful clinical information was swamped. It is well known that adding extra, noisy or irrelevant inputs to a classifier can reduce performance, and this may be why the clinical data alone is better in some cases. Other ways of combining this information would be a good avenue for future work.
- Overall Novelty (R3, R5) While we acknowledge that our network architectures are not novel, we highlight the novelty of combining multi-task learning and image synthesis for stroke prognostication.
- Training (R3, R4) Clinical information, when used, was concatenated with the stage 2 network output (class probability), forming the logistic regression inputs. All of our models were trained, independently, from scratch. Models using synthetic CTP maps never saw acquired CTP maps. Deeper networks for NCCT+CTA were tried in our initial architecture exploration, but performed worse, probably due to limited data. More training details and ROC statistics will be shown in our final version.
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.
This is a quite interesting paper which demonstrates synthesizing CT perfusion maps from NCCT and CTA could benefit the outcome prediction in large vessel stroke prognostication in terms of mRS score <=2 or >2 as a binary classification problem. The rebuttal has addressed the major concerns and the paper has interesting viewpoints to be accepted by MICCAI. I hope authors address the low-quality and spatial resolution of predicted CTP maps in the Fig. 5 and show them in pseudo colormaps as shown in commercial software for visualization.
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.
Overall, this is a good application-driven paper with clear motivation, and I think the authors’ responses could address the major concerns raised by the reviewers. Thus, I tend to side with most reviewers to accept the paper.
A key limitation is that the performance gains look marginal compared with other synthesis-based methods. Also, the authors are suggested to rephrase the description of their technical contribution from the application aspect, considering that down-stream task-guided synthesis in a multi-task learning setting is not new in the broader community.
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.
This paper aims to address stroke prognosis prediction based on methods of image synthesis. The paper is well-motivated and well-written. The task is interesting and less frequently explored in MICCAI. The rebuttal clarifies main objectives, method specifics and experimental details. It can be seen that the concerns of R3 are partially tackles. Overall, meta-reviewer thinks this is a well organized paper with decent quality, and would like to follow majority of reviewer ratings for a positive recommendation.