Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Automated alveolar cleft defect restoration from cone beam computed tomography (CBCT) remains a challenging task, considering large morphological variations due to inter-subject abnormal maxilla development processes and a small cohort of clinical data. Existing works relied on rigid or deformable registration to borrow bony tissues from an unaffected side or a template for bony tissue filling. However, they lack harmony with the surrounding irregular maxilla structures and are limited when faced with bilateral defects. In this paper, we present a stochastic anomaly simulation algorithm for defected CBCT generation, combating limited clinical data and burdensome volumetric image annotation. By respecting the facial fusion process, the proposed anomaly simulation algorithm enables plausible data generation and relieves gaps from clinical data. We propose a weakly supervised volumetric inpainting framework for cleft defect restoration and maxilla completion, taking advantage of anomaly simulation-based data generation and the recent success of deep image inpainting techniques. Extensive experimental results demonstrate that our approach effectively restores defected CBCTs with performance gains over state-of-the-art methods.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2024/paper/1091_paper.pdf

SharedIt Link: pending

SpringerLink (DOI): pending

Supplementary Material: https://papers.miccai.org/miccai-2024/supp/1091_supp.pdf

Link to the Code Repository

https://github.com/Code-11342/SAS-Restorer.git

Link to the Dataset(s)

N/A

BibTex

@InProceedings{Guo_Stochastic_MICCAI2024,
        author = { Guo, Yixiao and Pei, Yuru and Chen, Si and Zhou, Zhi-bo and Xu, Tianmin and Zha, Hongbin},
        title = { { Stochastic Anomaly Simulation for Maxilla Completion from Cone-Beam Computed Tomography } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2024},
        year = {2024},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15008},
        month = {October},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    This paper proposed a simulation algorithm to generate CBCT with alveolar cleft defect from normal CBCT. Additionally, the training pipeline with mixed data (clinical defected CT and simulated defected CT) and the symmetric constraint are proposed for defect restoration and maxilla completion.

  • 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 problem is interesting.
    2. The proposed method can improve the performance.
  • 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 motivation of Eq. 1 is not clear. In other words, why is such a simulation reasonable?
    2. Eq. 2, is V_r in the first term the restoration of each V (clinical data)? V_r is used in the first paragraph if Sec. 2.2 to represent the restoration of the simulated V_s.
    3. Dataset. Are the clinical data collected with bounding box annotations?
    4. Table 1, with simulated data, the improvements in clinical and simulated data are not consistent (limited in clinical data, significant in simulated data). This means the simulated data indeed have a considerable gap with the clinical data. Hence, the reasonableness of the proposed simulation process would be a concern.
    5. The method is limited to the augmentation of CBCT with alveolar cleft defect, and the authors should discuss the feasibility of the proposed simulation to be adapted to other diseases or organs.
  • Please rate the clarity and organization of this paper

    Satisfactory

  • 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 does not mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • 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

    See weaknesses.

  • 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?

    Limited improvements using simulated data and concerns about the feasibility of the proposed simulation.

  • 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

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

  • [Post rebuttal] Please justify your decision

    The authors addessed my concerns in the rebuttal.



Review #2

  • Please describe the contribution of the paper

    Authors present a stochastic anomaly simulation method for defected cone beam computed tomography (CBCT) generation respecting abnormal maxilla developments. They introduce a volumetric inpainting framework with adversarial restoration learning and craniofacial symmetry for harmonic volumetric restoration of missing bony tissues. The proposed technique is evaluated on a clinical defected CBCT dataset and compared with state-of-the-art methods.

  • 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.

    Experimental results show a good performance in both:

    • virtual maxilla completion and cleft defect volume estimation (DSC, AHD and MSD)
    • restoration accuracy (PSNR, SSIM and NCC), when compared with state-of-the-art methods (rigid mirror registration, Demons, SyN, VM , DAE, U-Net, GII).
  • 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.
    • Information about computational time is not presented.
    • Computing system (hardware) used in experiments is not described.
  • Please rate the clarity and organization of this paper

    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 submission does not mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • Do you have any additional comments regarding the paper’s reproducibility?
    • The code is not available, but enough details to reproduce it are given. -The dataset used by authors is described but it is not publicly available.
    • A description of the computing infrastructure used (hardware) is not presented.
  • 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

    It would be interesting that authors:

    • Give some information about the computational time required by the proposed method.
    • Give a reference (and/or formulas) for Dice similarity coefficient (DSC), the average Hausdorff distance (AHD), and the mean squared deviation (MSD).
    • Include a description of the computing infrastructure used.
    • Summarize the research limitations and future research directions.
  • 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 topic of the paper is relevant and interesting to the MICCAI community.
    • Although some information is missing in the results section, the results are promising and the conclusions are supported by the results.
  • Reviewer confidence

    Very confident (4)

  • [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 addressed my comments and the comments of the other reviewers, so it is recommended that this paper be accepted.



Review #3

  • Please describe the contribution of the paper

    This paper proposed the way to simulating to overcome limited number of dataset for the specific clinical application of alveolar cleft defect. They proposed the way to generate defected region from normal CBCTs, by computing stochastic skeleton (by introducing stochastic random variable for each path points). For the clinical test case, they created a restored CBCTs by combining simulated model and find the defected region with these two volumes and symmetric constraint.

  • 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 is pretty well written paper and easy to understand. The main technical novelty is in the way to generate defected cases from normal CBCTs by simulation to overcome the limited data size of the rare disease, which looks convincing, and clinically meaningful. They use this data to generated the restored CBCT to predict the defect region of the original clinical case. They compared traditional deformation based methods and some deep networks to show the better performance.

  • 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.

    The experiments are overally weak. Limited number of dataset can be understandable, but the main weakness is the comparison with other state-of-the-art algorithms. The authors compared with limited number of deep learning algorithms and from the table 1, it is hard to say that the suggested algorithm is statistically superior than U-Net_SD considering the very small test size to apply. In addition, they mentioned about GAN or autoregressive models, but there was no quantitative comparison with their proposed model.

  • 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 submission does not provide sufficient information for reproducibility.

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

    No

  • 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

    It reads well but some parameters are not well described for the reproducibility. For example, what is the value of mu_1 and mu_2 in your experiment? The skeleton points are generated until it reaches s_u? (Is it guaranteed?) This kind of complete details would be helpful for readers.

  • 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 idea to generate defected CBCTs by simulation sound convincing, even though experiments are pretty limited.

  • 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



Author Feedback

We thank the reviewers for their efforts in reviewing our paper and their constructive comments. @R3: Comparison with SOTA. We have compared with the deep learning-based methods, including VM [4], DAE [14], U-Net [6], and GII [23], as shown in Table 1 and Fig. 4. U-Net_SD [6] is trained using paired defected CBCTs and cleft defect masks of D_s generated by the proposed SAS algorithm. Our approach has shown performance gains of 0.02 (DSC) and 0.01 mm (AHD) over the U-Net_SD. We have compared with the GAN-based method of the GII [24]. The GII is suitable for filling the masked region, but it does not address the morphological variations of irregular cleft defects. Our approach outperforms GII by 0.05 and 0.10 regarding the DSC on D_c and D_s (Table 1). @R1, R4: Limitations and future research. The proposed SAS algorithm relied on iterative skeleton tracing and dilation, where the progressively generated fat skeleton was used to guide the defected CBCT generation. The proposed SAS algorithm is specific to the 3D craniofacial fusion process involving the primary or secondary palate. The extension of the SAS algorithm to adapt to diversified abnormal developments of other diseases or organs deserves further study. Another limitation is that the proposed restoration model produces missing bony tissues conditioned on the input defected CBCTs. Considering patient-specific bone resorption after the grafting procedure, we would further investigate time-varying restoration learning in a variety of clinical applications. @R4: Improvements in clinical and simulated data. When given simulated data for supervised learning, the model is feasible to improve restoration performance with a larger margin on D_s than D_c over the model learned from limited clinical defected CBCTs via adversarial restoration learning. We think the reason is that the testing simulated defected data in D_s bear the same distribution as the training data. Ablation study indicated that simulated data enhances the generalization capacity for cleft defect estimation. The SD also improves performances of existing DAE and U-Net (Table 1). @R4: Motivation of Eq. 1. Confronted with a small cohort of defected CBCT scans in the clinical study, we present the SAS algorithm to simulate the inverse facial fusion process. The SAS algorithm relies on iterative skeleton tracing and dilation for diversified defected CBCT generation under the Tessier system of orofacial clefting. Without loss of generality, we start skeleton-tracing from the bounding plane of the cleft defect. Eq. 1 defines the skeleton-growing vector q_i in the i-th step. We use a randomly perturbed vector a and the vector towards s_u to update the growing vector for diversified skeletons and simulated cleft defects (Sec. 2.1). @R1, R3, R4: Code availability. We would make the code available after acceptance. We clarify the concerns on implemental details as follows: @R1: Computational time and hardware. We evaluate the proposed model on a PC with an NVIDIA GeForce RTX 3090 GPU. The training and online inference take approximately 140 hours and 0.45 seconds, respectively. As to the metric formulas, DSC (X,Y)=(2|X∩Y|)/(|X|+|Y|). |X| and |Y| refer to the cardinality of sets X and Y, respectively. AHD(X,Y)=1/2(1/|X|\sum_{a\in X}\min_{b\in Y} |a-b|+1/|Y|\sum_{b\in Y}\min_{a\in X}|b-a|). MSD(X,Y)=1/2(1/|X|\sum_{a\in X}\min_{b\in Y}|a-b|+1/|Y|\sum_{b\in Y}\min_{a\in X} |b-a|). We would provide a reference to the metrics. @R3: \mu_1 and \mu_2 are set to 0.15 and 0.05 (Sec. 3). The skeleton tracing terminates when it reaches a predefined height or the upper bounding plane (Sec. 2.1). As the reviewer has pointed out, the simulated skeleton is not guaranteed to end up in s_u as Eq. 1. @R4: We use V_r to represent the restored volume in both supervised learning from simulated data and adversarial restoration learning from clinical data. The clinical data have bounding box annotations of cleft defect (Sec. 2.2).



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’

    The task is interesting, and the idea of simulating the data on CBCTs is also intriguing. However, as the reviewers have pointed out, the dataset is quite limited. Despite this, all reviewers have given a positive score; therefore, I recommend “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).

    The task is interesting, and the idea of simulating the data on CBCTs is also intriguing. However, as the reviewers have pointed out, the dataset is quite limited. Despite this, all reviewers have given a positive score; therefore, I recommend “accept”.



Meta-review #2

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

    Reject

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

    My main concern is the performance of the proposed restoration framework.

    • In Table 1, the fair comparison with the proposed method should be the U-Net_{SD} which also utilizes the simulated data. The improvement over U-Net_{SD} is much smaller than the std of the prediction. With some tuning of the U-Net architecture (residual block/#filters/#depth), the U-Net model may achieve better performance. It is unclear what is the #parameters and the runtime between simple U-Net and the proposed method. The incremental improvement may come from these hyperparameters.
    • The test set of real data only has 10 volumes, which makes the test result statistically less significant.
  • 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).

    My main concern is the performance of the proposed restoration framework.

    • In Table 1, the fair comparison with the proposed method should be the U-Net_{SD} which also utilizes the simulated data. The improvement over U-Net_{SD} is much smaller than the std of the prediction. With some tuning of the U-Net architecture (residual block/#filters/#depth), the U-Net model may achieve better performance. It is unclear what is the #parameters and the runtime between simple U-Net and the proposed method. The incremental improvement may come from these hyperparameters.
    • The test set of real data only has 10 volumes, which makes the test result statistically less significant.



Meta-review #3

  • 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



back to top