Abstract

Chronic subdural hematoma (cSDH) is a common neurological condition characterized by the accumulation of blood between the brain and the dura mater. This accumulation of blood can exert pressure on the brain, potentially leading to fatal outcomes. Treatment options for cSDH are limited to invasive surgery or non-invasive management. Traditionally, the midline shift, hand-measured by experts from an ideal sagittal plane, and the hematoma volume have been the primary metrics for quantifying and analyzing cSDH. However, these approaches do not quantify the local 3D brain deformation caused by cSDH. We propose a novel method using anatomy-aware unsupervised diffeomorphic pseudo-healthy synthesis to generate brain deformation fields. The deformation fields derived from this process are utilized to extract biomarkers that quantify the shift in the brain due to cSDH. We use CT scans of 121 patients for training and validation of our method and find that our metrics allow the identification of patients who require surgery. Our results indicate that automatically obtained brain deformation fields might contain prognostic value for personalized cSDH treatment.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2024/paper/1955_paper.pdf

SharedIt Link: pending

SpringerLink (DOI): pending

Supplementary Material: N/A

Link to the Code Repository

https://github.com/MIAGroupUT/Brain-Shift

Link to the Dataset(s)

N/A

BibTex

@InProceedings{Imr_BrainShift_MICCAI2024,
        author = { Imre, Baris and Thibeau-Sutre, Elina and Reimer, Jorieke and Kho, Kuan and Wolterink, Jelmer M.},
        title = { { Brain-Shift: Unsupervised Pseudo-Healthy Brain Synthesis for Novel Biomarker Extraction in Chronic Subdural Hematoma } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2024},
        year = {2024},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15002},
        month = {October},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The paper proposes a novel unsupervised framework for image distortion estimation. Based on this system, the paper also proposes three deformation-based metrics for disease severity estimation. Classification experiments are conducted to show that these metrics can better capture the severity of cSDH compared with conventional measurements.

  • 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. The proposed framework is capable for multi-tasks, including segmentation and distortion correction.
    2. The paper proposes multiple novel unsupervised loss formulations such as symmetry loss and ventricle loss.
    3. Experiments show that the computed deformation-based metrics can perform better than conventional metrics to some extent.
  • 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. The proposed model is not end-to-end trainable, the previous mistakes would accumulate across the framework. However, I think, for the segmentation module and deformation module could be combined and optimized jointly.
    2. The proposed loss functions might cause unnecessary deformation. Such as ventricle symmetry loss, it might cause the unrealistic ventricle shapes as shown in Fig3.
    3. The experimental section is not well-established. Only simplistic analysis is presented. The effects of different loss functions are not investigated and discussed.
    4. The performance improvement of the proposed metrics is not significant compared with the conventional metrics. And the evaluation is only performed on one small-size in-house dataset.
  • Please rate the clarity and organization of this paper

    Very Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Do you have any additional comments regarding the paper’s reproducibility?

    N/A

  • Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review. Pay specific attention to the different assessment criteria for the different paper categories (MIC, CAI, Clinical Translation of Methodology, Health Equity): https://conferences.miccai.org/2024/en/REVIEWER-GUIDELINES.html
    1. An ablation study of different loss functions is preferred.
    2. An evaluation on public dataset would be better for generalization estimation of the method.
    3. Since the framework introduces multiple symmetrical losses, the deformation would also change the geometry of the healthy half for unilateral cases. Which might cause the proposed metric’s performance get worse as shown in Fig 4 for total S column.
    4. A more detailed and quantitative explanation or the improvement for proposed metrics are necessary. Such as a T-test can help other researchers to understand the scale of improvement for each metric in each group.
  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making

    Weak Reject — could be rejected, dependent on rebuttal (3)

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    This paper proposes a novel framework to estimate the severity of the cSDH by combining the segmentation and translation methods. However, the experiment is in sufficient, the small-size dataset used is in-house, and the lack of ablation study and quantitative results. Also, the proposed loss functions are heuristic. So, I think the drawbacks overshadow the novelty behind the key idea of this paper.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [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 #2

  • Please describe the contribution of the paper

    This work proposes an unsupervised diffeomorphic pseudo-healthy CT scans using symmetry losses for personalized treatment prediction. The experiment showed that the imaging biomarker discovered by the proposed method could be more sensitive particularly for bilateral chronic subdural hemorrhage patients.

  • 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.
    • The proposed method is technically sounding and well described.
    • The paper is well written.
    • The novel deformation biomarker is shown to be more sensitive to bilateral cSDH cases over traditional biomarkers.
  • 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.
    • Relatively small data size compared to other works for quantifying MLS.
    • The technical novelty is not clearly highlighted against prior work.
  • Please rate the clarity and organization of this paper

    Very Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Do you have any additional comments regarding the paper’s reproducibility?

    N/A

  • Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review. Pay specific attention to the different assessment criteria for the different paper categories (MIC, CAI, Clinical Translation of Methodology, Health Equity): https://conferences.miccai.org/2024/en/REVIEWER-GUIDELINES.html
    • Eq (5): How were these lambda values optimized?
    • Sect 3.3:”… MLS poorly separates patients with bilateral cSDH.” It would be nice if an example image can be provided for this statement; How were MLS obtained?
    • Table 1: it would be useful to comment on why sum is more accurate than max for the bilateral cases?
    • Would it be possible to comment how hemorrhage would affect the initial alignment?
    • Discussion on how the model would work for cases with acute hem subtypes such as IPH or SAH would be useful
  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making

    Weak Accept — could be accepted, dependent on rebuttal (4)

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
    • The paper is well organized and written
    • The proposed method would be interesting to the MICCAI audience
    • The validation result highlights the potential benefit of the method well.
  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [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 #3

  • Please describe the contribution of the paper

    The authors introduce an image-to-image method for synthetic generation of pseudo-healthy brain CT scans from real scans of patients with cSDH, which is trained using unsupervised losses. The synthetic maps, together with the real CT scans, are then used to generate deformation fields over the brain volume, which enable the computation of three novel biomarkers of brain deformation due to the exerted pressure from the cSDH. These biomarkers, together with two additional classic biomarkers formerly described in the literature (midline shift and hematoma volume), are then used to fit a logistic regression model to classify patients depending on whether they received invasive surgery or not. The use of the novel biomarkers leads to improved discrimination ability by the model, especially for unilateral cSDH.

  • 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.
    • The evaluation of conditions associated with brain mass-effect diseases is well-structured and presented.
    • It introduces a game-changing approach that challenges the traditional use of simple metrics and arbitrary thresholds from decades ago.
    • A clever workaround is devised to create a reference healthy dataset from pathological data (baseline CT scans), enabling the inference of deformation fields.
    • Limitations are correctly addressed
  • 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.
    • Evaluation bias is a significant concern. The ground-truth is subjective, as acknowledged in the limitations. Moreover, the data is sourced from a single center, amplifying bias. For instance, relying solely on the subjective perspective of treating physicians without an evidence-based decision-making algorithm poses challenges. It would be beneficial if the authors provided at least a minimal explanation of why patients are treated or not. Typically, such decisions are driven by clinical factors. Did the authors compare whether clinical indicators (e.g., Glasgow Coma Scale, early deterioration, focal neurological deficits) perform comparably to deformation markers? It’s crucial to remember that we treat patients, not CT scans, and their clinical status should always be considered, especially when evaluating aggressive treatments like those used for chronic subdural hematoma.
    • A hold-out test evaluation would be valuable to ascertain the absence of clear overfitting. However, the small sample size precludes such a design. -The approach seems somewhat lacking in justification. While the generated biomarkers indeed enhance classification accuracy, the overall solution appears complex. The deformation maps present intriguing intermediate results and hold promise as a valuable tool. Yet, it’s worth exploring more sophisticated classification models that leverage deformation maps, rather than directly using aggregates (maximum, average and total) of the complex information derived from this work. Alternatively, the authors could delve into potential avenues for further exploitation of the deformation maps in other fields.
  • Please rate the clarity and organization of this paper

    Excellent

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission has provided an anonymized link to the source code, dataset, or any other dependencies.

  • Do you have any additional comments regarding the paper’s reproducibility?

    Code is provided. Data is from a private dataset. All training parameters are detailed in the paper

  • Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review. Pay specific attention to the different assessment criteria for the different paper categories (MIC, CAI, Clinical Translation of Methodology, Health Equity): https://conferences.miccai.org/2024/en/REVIEWER-GUIDELINES.html

    I would like to congratulate the authors on what I consider a well-presented article, with multiple methods specifically designed to solve a task-specific problem, considering and taking into account the nature of the problem itself, and presenting smart solutions to overcome significant problems such as the lack of healthy image pairs.

    Having said that, I would like to make a few comments:

    • The number of patients with bilateral and unilateral patients should be reported.
    • It would be nice to see the error rate on the extraction of the deformation maps (the authors comment on the accumulated errors as a limitation, but fail to provide what is the prevalence of these errors in their sample).
    • Is there any attempt to implement an image classification method as baseline from the original CT scan? Could this be experimented with in this cohort to have an additional baseline of a much simpler approach? If the results were worse, this could help justify the approach taken in this paper. Otherwise, it could be argued that albeit thoughtful and well-made, the synthetic pseudo-healthy scan generation could be an unnecessary overhead, unless the authors justify if and how the deformation maps could add value in other ways (perhaps these maps could be used as a means of explainability?).
    • What is the reason for limiting the implemented classification head to a logistic regression model? I understand that you could do that to maintain a certain transparency of the model and better grasp how the features are used for final classification, but that is not expanded on in the results or discussion. Perhaps the generated deformation map together with the original scan could be fed to a classification network. This could be a nice idea to test. If I understand correctly, by reducing the biomarkers to a set of aggregated features (i.e. a single number) you are losing the spatial component of the deformation map. Perhaps an adequate model could learn how to leverage it for more precise classification.
    • Could diffusion models present a more attractive alternative to the pseudo-healthy synthetic generation of the brain CT scans? I would love to hear the author’s take on this point in the discussion or in the rebuttal response.
  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making

    Accept — should be accepted, independent of rebuttal (5)

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The authors present an interesting and rigorous approach of an advanced image analysis pipeline for cSDH treatment necessity prediction. I believe this is still a work-in-progress and the authors need to work on further justifying the particularities of their solution compared to simpler approaches (i.e. the deformation maps generated from their model could potentially be exploited further).

  • Reviewer confidence

    Somewhat confident (2)

  • [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed

    Accept — should be accepted, independent of rebuttal (5)

  • [Post rebuttal] Please justify your decision

    The authors presented a rebuttal letter replying to concerns raised by reviewers. Although there are some limitations/biases that should be addressed, I believe that the current work is a proof-of-concept study that can be developed in a way that can be extremely useful in clinical practice and should be further explored. I maintain my original rating 




Author Feedback

We thank the reviewers for their kind words and valuable feedback. Below, we address the key points raised.

R4 raised concerns about the bias introduced to our dataset since we source data from one center. We are aware of this limitation and are currently setting up a larger multi-center consortium to validate our results. With this initiative, we also aim to obtain additional objective target outcomes in consultation with clinical partners. Indeed, our current labels (treatment or not) might be influenced by patient-specific factors that cannot be extracted from CT. Nevertheless, we expect that our unsupervised feature extraction approach might provide valuable information for novel targets.

R4 questioned the use of aggregate measures derived from the deformation fields. Indeed, it is likely that the ‘raw’ deformation fields contain additional information that might be extracted using, e.g., a separate CNN. However, we chose not to train a model directly from the deformation fields due to the limited size of our dataset. However, we agree with R4 that exploring more complex classification models using deformation maps is a valuable future study.

We agree with R3 that an evaluation of our method in a public benchmark would provide a better way to assess the generalization of the method. Unfortunately, to our knowledge, there are no public datasets of 3D CT images of patients with cSDH, for neither surgery prediction nor segmentation. We have contacted authors of related segmentation works but found that they were reluctant to share data.

R3 noted that our model is not trained end-to-end. There are practical benefits to the current modular approach: By considering segmentation a task that is independent of pseudo-healthy brain synthesis, we are able to easily swap out trained segmentation models. On the other hand, end-to-end training of the segmentation, pseudo-healthy brain synthesis, and even the classification task might improve performance in all stages. Hence, end-to-end training is interesting for future work. We will add this to the Discussion.

Following questions from R1 and R3, we would like to further clarify the terms in the loss function (Eq. 5). Each term in the loss function aims to guide the deformation to a more uniform and anatomically plausible form. R3 noted the ventricle symmetry loss might cause an unrealistic ventricle shape. This is partially true, as the final ventricle system in Fig. 3 is not perfect, and the effects of the “forced symmetry” can be seen. However, we found that the inclusion of the ventricle loss in our model had a positive effect. Most of the compression due to cSDH occurs within the ventricles. By explicitly forcing this volume to be regained by the ventricle system, we prevent excessive deformation elsewhere in the brain. We will include this information in the Discussion. In future work, a detailed ablation study could help shed further light on the contribution of each loss term. The open-source implementation of our paper will include all the coefficients in Eq. 5, which are generally kept at 1.0. We used these to balance the spectrum of our loss functions (keeping them between 0 and 1 as much as possible), rather than to rank them in importance.

Following the comments of R1, we’d like to clarify our MLS measurement protocol. The MLS was measured following the clinical protocol from our clinical center and validated through in-person consultations with our clinical partners. We will include this in the final version of our paper.

R1 asked about how our model would work with regard to different types of conditions. This model is developed for cSDH so will likely work for acute subdural or epidural hematomas. ⁤⁤However, hemorrhagic strokes like SAH or IPH have different characteristics. ⁤⁤Applying our findings to these hemorrhages would be interesting, but would also require caution, as their diagnosis and decision-making processes differ significantly from cSDH.




Meta-Review

Meta-review #1

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A

  • 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).

    N/A



Meta-review #2

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    The authors properly acknowledge and explain the limitations of their method and experimental results. The minor modifications to the paper will improve it.

  • 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).

    The authors properly acknowledge and explain the limitations of their method and experimental results. The minor modifications to the paper will improve it.



back to top