List of Papers Browse by Subject Areas Author List
Abstract
Automated segmentation of stroke lesions on non-contrast CT (NCCT) images is essential for efficient diagnosis of stroke patients. Although diffusion probabilistic models have shown promising advancements across various fields, their application to medical imaging exposes limitations due to the use of conventional isotropic Gaussian noise. Isotropic Gaussian noise overlooks the structural information and strong voxel dependencies in medical images. In this paper, a novel framework employing synchronous diffusion processes on image-labels is introduced, combined with a sampling strategy for anisotropic noise, to improve stroke lesion segmentation performance on NCCT. Our method acknowledges the significance of anatomical information during diffusion, contrasting with the traditional diffusion processes that assume isotropic Gaussian noise added to voxels independently. By integrating correlations among image voxels within specific anatomical regions into the denoising process, our approach enhances the robustness of neural networks, resulting in improved accuracy in stroke lesion segmentation.
The proposed method has been evaluated on two datasets where experimental results demonstrate the capability of the proposed method to accurately segment ischemic infarcts on NCCT images. Furthermore, comparative analysis against state-of-the-art models, including U-net, transformer, and DPM-based segmentation methods, highlights the advantages of our method in terms of segmentation metrics. The code is publicly available at https://github.com/zhangjianhai/SADPM.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2024/paper/1208_paper.pdf
SharedIt Link: https://rdcu.be/dV1PN
SpringerLink (DOI): https://doi.org/10.1007/978-3-031-72069-7_41
Supplementary Material: N/A
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{Zha_Synchronous_MICCAI2024,
author = { Zhang, Jianhai and Wan, Tonghua and MacDonald, M. Ethan and Menon, Bijoy K. and Qiu, Wu and Ganesh, Aravind},
title = { { Synchronous Image-Label Diffusion with Anisotropic Noise for Stroke Lesion Segmentation on Non-contrast CT } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2024},
year = {2024},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15002},
month = {October},
page = {433 -- 443}
}
Reviews
Review #1
- Please describe the contribution of the paper
this paper proposes a novel method that samples anisotropic gaussian noise for DDPMs.
- 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.
comparison to state of the art methods shows superior performance of the proposed method strong quantitative metrics. good theory and explanations of the algorithms.
- 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.
i disagree with the motivations listed in this paper. it is true that medical images have less variation in shape / size / illumination / etc than large scale datasets such as ImageNet or LAION-5B. However - there are many other smaller specific datasets - such as FFHQ - where the exact same properties would hold. Please see comments to author for detailed feedback on this.
- 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 provide sufficient information for 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
[major]
-
in abstract - “which overlooks the structural information and strong voxel dependen- cies in medical images “ - don’t natural images also have structure and pixel dependencies? i think this is a wrong way to frame the problem. spatially dependent information is contained in all images (except for pure noise).
-
in abstract - “Our method acknowledges the sig- nificance of anatomical information during diffusion “ - i agree that medical images have a more structured spatial appearance (less variation) due to a standardized acquisition and fairly narrow variation in anatomy. i disagree that this is not possible to achieve for natural images. if i were to take a FFHQ dataset (of faces) - it has exact same properties like you describe.
-
Re: “However, most of the applications are demonstrated in the domain of natural images, such as: portraits, paysage, and camera surveillance, where isotropic Gaussian noise is usually assumed and used in the denoising process. By contrast, medical images possess unique charac- teristics, stable image structure, immobilized intensities and texture etc., that make isotropic noise unsuitable for segmentation tasks “ i think this is a strong mischaracterization of DDPMs. If you read the original paper by sohl-dickstein (“Deep unsupervised learning using nonequilibrium thermodynamics”, 2015) - you will note that the gaussian nature of DDPM like processes is not modelled after the ‘images have gaussian properties’. Secondly, you can frame the same statement about FFHQ dataset (or any other dataset of standardized single class objects).
-
Consider changing abstract, introduction and section 2.2 with an updated motivation for the paper.
-
since the main contribution of this paper is the formulation of sampling from anisotropic noise - i would urge the authors to expand section 2.3 since it is currently very difficult to read in such a short paper. To do this - reduce the basic facts about DDPM in section 2.1. - most of these formulations here come from standard DDPM theory (except for dual image sampling). You can simply refer to SDPM paper and make this section 2.1. at most one-two paragraphs (you state clearly that SADPM is just an extension of SDPM with anisotropic noise anyway). Also - section 2.2 - is kind of unnecessary here - most information is repeated from introduction.
[minor] a - We extent the diffusion model - > We extenD the diffusion model
b - with the same Gasussian -> Gaussian
c - isotropic Gassusian noise -> Gaussian
-
- 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?
i disagree with motivations of the authors which set out to build anisotropic gaussian sampling for DDPMs. If this is reframed - i think the paper has merit as it shows superior results AND the actual process of sampling anisotropic gaussian is interesting and novel in itself.
- 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 paper presents a fully probabilistic inference framework for stroke lesion segmentation on NCCT (non-contrast CT) images based on a synchronous image-label denoising diffusion model with anisotropic Gaussian noise. Experiments compared with recent state-of-the-art showcase the advantages of the proposed approach.
- 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 overall writing and presentation is clear and easy to follow.
- The intuition for using anisotropic gaussian seems reasonable and interesting.
- Experiments are extensive, including comparisons to different SOTA architectures. The ablations on anisotropic gaussian indicates the superiority of the proposed method.
- 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.
- It would be more insightful if the authors include the computational costs and related comparisons of their approach v.s. Conv Nets or DDPMs.
- Minor comment: the authors did not mention open access to source code nor its reproducibility.
- 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?
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
Please refer to the weakness section above.
- 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?
Please see the strengths and weakness sections.
- 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
Weak Accept — could be accepted, dependent on rebuttal (4)
- [Post rebuttal] Please justify your decision
Thank you for you response. The rebuttal clarified my concerns to some extent, therefore I tend to remain my positive score as leaning to accept.
Review #3
- Please describe the contribution of the paper
The authors introduced a novel framework that employs a synchronous diffusion process on image labels, and by combine with an anisotropic noise sampling strategy, they have improved stroke lesion segmentation performance on NCCT data.
- 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 methodology is described as a series of formulations, coupled with some pseudocode, making it quite clear and easy to follow. The presented results demonstrate the improvements brought by the newly introduced methods both quantitively and qualitatively. However, it would be even better if some comments and evidence on a comparison of computation efficiency are provided. This would give a more complete picture of how applicable the method is to clinical settings, where computation resources are often scarce.
- 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 description of all the key components of the propose methods is clear, however, it would be beneficial to the readers if a high-level overview of the entire pipeline is also presented.
- 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?
Links to the source code and the private dataset (some descrition and basic stastics) would be nice for the sake of reproducibility.
- 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
Overall, the presented method is novel and clearly described for the most part. However, a high-level overview of the entire pipeline (e.g. some flowchart) would be nice. The experiment setups and results are nicely demonstrated. More qualitatively results would be great, these can be included in the appendix for example. Lastly link to the source code and the private dataset would benefit the reproducibility.
- 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 overall presentation of the paper is quite clear and the introduced method is novel and would contribute to field in question. In addition, the results have clearly shown the improvements when compared to other state-of-the-art methods.
- 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 appreciate the thoughtful feedback from the reviewers. We will correct errors and clarify the major concerns from the reviewers in the final version.
1) Reproducibility: We will make code publicly available after acceptance. 2) Computational costs: Our fast-sampling algorithm of anisotropic noise can greatly reduce the computational costs. In the experiments, our method based on synchronous image-label diffusion model framework is much faster than the comparison methods of SegDiff and MedSegDiff using conditional diffusion models. The related analyses will be extended in the future work. 3) Motivation (#3): We agree with the reviewer that natural images have structure and pixel dependencies, and spatially dependent information is contained in all images. We acknowledge that datasets like FFHQ share structured spatial properties with medical images. We clarify our motivations to emphasize the specific challenges and characteristics of medical imaging while recognizing that similar structured spatial dependencies can exist in certain natural image datasets. The phrase “which overlooks the structural information” in abstract refers to the standard Gaussian noise (i.e., multivariate standard normal N(0,I)), not the images or diffusion models. We would like to emphasize that the traditional isotropic noise is a pure stochastic noise, where the variables are independent and covariances among different variates are zeros. Our anisotropic Gaussian noise is sampled from a distribution of N(0, Sigma)， where Sigma is not an identity matrix, the diagonal entries are not ones and the off-diagonal entries are not zeros, indicating spatial dependencies. This allows for pixel dependencies and preserves image structure. It is consistent with the reviewer’s opinions. We will reframe our motivations throughout the paper and avoid misunderstandings and ensure clarity.
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’
This paper presents a fully probabilistic inference framework for stroke lesion segmentation on NCCT (non-contrast CT) images based on a synchronous image-label denoising diffusion model with anisotropic Gaussian noise. The reviewers are generally in favor of the paper. The authors shall carefully address the remaining concerns in their final version.
- 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).
This paper presents a fully probabilistic inference framework for stroke lesion segmentation on NCCT (non-contrast CT) images based on a synchronous image-label denoising diffusion model with anisotropic Gaussian noise. The reviewers are generally in favor of the paper. The authors shall carefully address the remaining concerns in their final version.
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’
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