List of Papers Browse by Subject Areas Author List
Abstract
Efficient and accurate brain ventricle segmentation from clinical CT scans is critical for emergency surgeries like ventriculostomy. With the challenges in poor soft tissue contrast and a scarcity of well-annotated databases for clinical brain CTs, we introduce a novel uncertainty-aware ventricle segmentation technique without the need of CT segmentation ground truths by leveraging diffusion-model-based domain adaptation. Specifically, our method employs the diffusion Schrödinger Bridge and an attention recurrent residual U-Net to capitalize on unpaired CT and MRI scans to derive automatic CT segmentation from those of the MRIs, which are more accessible. Importantly, we propose an end-to-end, joint training framework of image translation and segmentation tasks, and demonstrate its benefit over training individual tasks separately. By comparing the proposed method against similar setups using two different GAN models for domain adaptation (CycleGAN and CUT), we also reveal the advantage of diffusion models towards improved segmentation and image translation quality. With a Dice score of 0.78±0.27, our proposed method outperformed the compared methods, including SynSeg-Net, while providing intuitive uncertainty measures to further facilitate quality control of the automatic segmentation outcomes. The code is available at: https://github.com/HealthX-Lab/DiffusionSynCTSeg.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2024/paper/1439_paper.pdf
SharedIt Link: pending
SpringerLink (DOI): pending
Supplementary Material: N/A
Link to the Code Repository
https://github.com/HealthX-Lab/DiffusionSynCTSeg
Link to the Dataset(s)
N/A
BibTex
@InProceedings{Tei_CTbased_MICCAI2024,
author = { Teimouri, Reihaneh and Kersten-Oertel, Marta and Xiao, Yiming},
title = { { CT-based brain ventricle segmentation via diffusion Schrödinger Bridge without target domain ground truths } },
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
Authors propose a diffusion Schrodinger Bridge and an attention recurrent residual U-Net segmentation model (E2E-UNSB+R2AUNet) and capitalize on unpaired CT and MRI scans to derive automatic CT segmentation from those of the MRIs. E2E-UNSB+R2AUNet essentially is an end-to-end joint training framework of image translation and segmentation tasks. Compare with two GAN models (CycleGAN and CUT), E2E-UNSB+R2AUNet also reveal the advantage of diffusion models towards improved segmentation and image translation quality.
- 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.
- Authors propose a novel insight for the CT-based brain ventricle segmentation, where the ‘precise’ labels of MRI could be utilized to the ‘lack of well-annotated’ public CT via style translation.
- 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 motivation of uncertainty assessment is not clear, since it might not make the contribution to the insight of translation and segmentation stages.
- Besides SynSeg-Net, it is suggested to compared with the other existing methods.
- 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 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
Authors propose and end-to-end unified method for style translation (MRI-to-CT) and CT-based brain ventricle segmentation. The insightful application (w/o ground truth-based style translation) is well in the community of brain ventricle segmentation. However, some issues should be considered in their future works as follows:
- The motivation of uncertainty assessment is not clear, since it might not make the contribution to the insight of translation and segmentation stages.
- Besides SynSeg-Net, it is suggested to compared with the other existing methods.
- 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?
Strengthen: Motivation of style translation. Weakness: Weak motivation of uncertainty assessment and unsufficient experiment.
- 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
N/A
- [Post rebuttal] Please justify your decision
N/A
Review #2
- Please describe the contribution of the paper
This paper presents a ventricle segmentation model using CT brain images. Major contribution of this work is that the method allows MRI-to-CT adaption and the model training leverages ventricle segmentation in MRI scans, without the need to label ground truths in CT.
- 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.
Main strength of this work is the use of unpaired neural Schrodinger Bridge to allow MRI-CT translation in a joint framework with segmentation task. End-to-end training and the joint loss function for both image translation and segmentation result in good performance in ventricle segmentation using CT.
- 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.
Weakness of this work is the experiment of performance comparison. Segmentation accuracy is based on the 27 sets paired MRI-CT data and the leading margin of the proposed method over the convention method using CUT is relatively small. A larger amount of testing data may help to provide more convincing results.
- 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 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
Authors of this paper are encouraged to perform thorough experiments for accuracy comparison with more SOTA methods using a larger amount of testing data. In particular, it is interesting to know whether the proposed method outperforms others using CT ground truths for training.
- 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?
Idea is new and proposed method is sound. Although more experiments are anticipated, this paper deserves a publication and presentation in MICCAI.
- 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
N/A
- [Post rebuttal] Please justify your decision
N/A
Review #3
- Please describe the contribution of the paper
Authors have presented a novel segmentation approach for CT segmentation without labels. In this approach unpaired CT and MRI scans / labels are being used in end-to-end and two-stage frameworks that includes synthetic CT and segmentation. They demonstrated that an end-to-end frame work with Unpaired Neural S Bridge for unpaired MRI-to-Ct combined with attention recurrent residual U-net can provide best segmentation results.
- 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.
Segmentation of soft tissue in CT scan due to limited accessibility of pixel-wise labels is always challenging. Synthetic segmentation without target labels has been introduced previously. Current study has shown that such frame work and leveraging UNSB and R2AUnet, can improve the segmentation of CT scan images, which can increase confidence in using CT segmentation tool in clinical practice.
Paper is well written and methodology and previous work were described clearly.
- 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.
while the training data is collected from multiple sources, test data belong to only one center. Do the authors believe similar performance is expected if different dataset was used?
- 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 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
The authors address CT segmentation, when labels from CT scan is not accessible. This is a very common situation specially in segmentation of soft tissue, where MRI labels are abundant, but CT is limited.
- 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
Strong Accept — must be accepted due to excellence (6)
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The authors address CT segmentation, when labels from CT scan is not accessible. This is a very common situation specially in segmentation of soft tissue, where MRI labels are abundant, but CT is limited.
Paper is well written and methodology and previous work were described clearly.
- 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
N/A
- [Post rebuttal] Please justify your decision
N/A
Author Feedback
We thank all the reviewers for their valuable feedback and respond to each of the comments below.
- Benefit of UNSB over CUT and GAN-based approaches (R3):
Through validation of the end-to-end approaches (the better setup), the superior segmentation results of the Unpaired Neural Schrödinger Bridge (UNSB) over GAN-based counterparts such as CycleGAN and CUT is confirmed via two-sided, paired-sample t-tests (p<0.05). The benefit of UNSB over CUT and CycleGAN is more evident when assessing segmentation in 3D (Dice=0.73, 0.69, and 0.56 for UNSB, CUT, and CycleGAN). In addition, with improved training stability, UNSB also offers a better SSIM metric than CUT for MRI-to-CT translation (~6% improvement), which also contributes to the segmentation outcome.
- Small test dataset and model adaptability (R1,R3):
Public paired MRI-CT datasets of the human head are rare, but are crucial for validating our proposed framework and the baselines. Here, we used the iDB dataset as our test set. We performed the evaluation both in a 2D slice-by-slice manner and in full 3D to take full advantage of the test dataset. As multi-centre data of unpaired MRI and CT scans were used for model training, the performance of the proposed method was shown to be great despite that no target-domain ground truths were used and the test set comes from another source. This showcased the robustness and adaptability of our model. Nonetheless, additional validation on multi-centre datasets would provide valuable insights into the model’s generalizability and performance consistency.
- Motivation of uncertainty estimation for segmentation (R4)
As target-domain ground truths were not used for training, uncertainty estimation for ventricle segmentation via Monte Carlo dropout helps enhance the reliability and interpretability for the deep learning model’s segmentation outcomes. With intuitive visual maps, the uncertainty map helps identify areas where the model is less confident. This is invaluable for clinical decision-making (e.g., planning ventriculostomy surgery). We also computed the correlations between model uncertainty and segmentation accuracy across different methods, further confirming the robustness of the proposed technique.
- Comparison with other SOTA models (R3,R4):
Our study uniquely addresses the challenge of CT brain ventricle segmentation in the absence of manual labels (partially due to poor soft tissue contrast and image quality), which is common in the clinical but often overlooked in the literature. So far, solutions to tackle this issue and particular application are limited. Typical SOTA methods use fully-labeled datasets. For instance, Huff et al. ( IJCARS, vol 14, 2019) achieved a high Dice score of 0.92 for the same task with fully labeled in-house datasets, but their method does not tackle the label scarcity issue that we intend to solve. Due to the obstacles in curating fully labeled CT brain dataset, it is difficult to compare our methods with supervised SOTA techniques. To validate the proposed method, we have compared two major SOTA methods, including CUT and CycleGANs with different variants to fully test and characterize our proposed framework, and showcased the superior performance of our method. We will further curate a larger dataset (with manual ground truths for CT scans) to fully validate our proposed framework with additional SOTA methods in our future studies.
Meta-Review
Meta-review not available, early accepted paper.