List of Papers Browse by Subject Areas Author List
Abstract
Multiphase CT angiography (mCTA) has become an important diagnostic tool for acute ischemic stroke (AIS), offering insights into occlusion sites and collateral circulation. However, its broader application is hindered by the need for specialized interpretation, contrasting with the intuitive nature of CT perfusion (CTP). In this work, we propose a novel diffusion based generative model to generate CTP-like perfusion maps, enhancing AIS diagnosis in resource-limited settings. Unlike traditional diffusion models that restore images by predicting the added noise, our approach uses a masked residual diffusion probabilistic model (MRDPM) to recover the residuals between the predicted and target image within brain regions of interests for more detailed generation. To target denoising efforts on relevant regions, noise is selectively added into the brain area only during diffusion. Furthermore, a Multi-scale Asymmetry Prior module and a Brain Region-Aware Network are proposed to incorporate anatomical prior information into the MRDPM to generate finer details while ensuring consistency. Experimental evaluations with 514 patient images demonstrate that our proposed method is able to generate high quality CTP-like perfusion maps, outperforming several other generative models regarding the metrics of MAE, LPIPS, SSIM, and PSNR. The code is publicly available at https://github.com/UniversalCAI/MRDPM-with-RAP.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2024/paper/1246_paper.pdf
SharedIt Link: https://rdcu.be/dV1Oj
SpringerLink (DOI): https://doi.org/10.1007/978-3-031-72069-7_26
Supplementary Material: N/A
Link to the Code Repository
https://github.com/UniversalCAI/MRDPM-with-RAP
Link to the Dataset(s)
N/A
BibTex
@InProceedings{Cai_Masked_MICCAI2024,
author = { Cai, Yuxin and Zhang, Jianhai and He, Lei and Ganesh, Aravind and Qiu, Wu},
title = { { Masked Residual Diffusion Probabilistic Model with Regional Asymmetry Prior for Generating Perfusion Maps from Multi-phase CTA } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2024},
year = {2024},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15002},
month = {October},
page = {270 -- 280}
}
Reviews
Review #1
- Please describe the contribution of the paper
Authors introduce a generative model capable of mapping from multiphase CT angiography images to CT perfusion images in the context of acute ischemic stroke detection. Their approach builds upon residual diffusion, additionally guided through region awareness (selective noise introduction based on segmentation), an “asymmetry prior” (emphasising potential stroke induced asymmetry) and additional semantic contextual feature extraction (further addressing the locality within the brain). They evaluate their work against SOTA methods and provide additional ablation studies.
- 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 address the important topic acute ischemic stroke detection in resource restricted context, where a well validated generative model potentially can offset limited specialised expertise. To the best of my knowledge, there is no prior work of using diffusion in the chosen context.
Localising and guiding diffusion through clinically relevant concepts is a promising idea, potentially capable of adding interpretability next to improved result-characteristics.
The work reasonably explains the influence, resp. role, of its components, and provides accompanying ablation studies.
Authors provide code for reproducibility.
- 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 diffusion (resp. residual diffusion) is an established and widely concept in the general deep-learning community by now, it is still limited within the medical context. The text relies on a working knowledge of it, limiting its audience in the wider community.
The work hints at clinical applicability but lacks key characteristics for this like timings and computational overhead.
Unclear what statistics where used, i.e. are the shown results averages over the dataset, or runs etc. While common, comparing models based on single instance is not a valid approach due to the stochastic nature of training.
- 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 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?
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
Combining diffusion with clinically relevant priors is promising direction. A higher abstraction of the concepts, i.e. addressing a bit more the intuition of the reader, potentially increases the appeal of the work to a wider audience.
- 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 work proposes a novel (in the applications context), promising approach for diffusion based image to image translation, with ample justification of the chosen components. While the evaluations seem to lack relevant characteristics and statistical rigor, utilising clinically relevant concepts for guiding diffusion is of potential of interest for the wider community.
- 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
N/A
- [Post rebuttal] Please justify your decision
N/A
Review #2
- Please describe the contribution of the paper
This paper presents a masked residual diffusion probabilistic model (MRDPM) incorporated with anatomical prior information, to recover the residuals between the predicted and target image within brain regions of interests for generation. Extensive experiments and ablations showcase the superiority 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.
- Code is provided and is promised to be publicly available.
- The experimental setups are reasonable and the ablations provided are insightful for readers to better understand the contributions from each proposed module.
- The Multi-scale Asymmetry Prior design seems interesting and reasonable.
- 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.
Minor comment: in Fig. 4, it might be easier for readers to observe with colored images instead of the grey-scale visualization for perfusion maps.
- 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 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?
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 see 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
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?
Please refer to 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
N/A
- [Post rebuttal] Please justify your decision
N/A
Review #3
- Please describe the contribution of the paper
The paper describes a novel diffusion model architecture to achieve high-quality multiphase CT Angiography (mCTA) from lower-quality images such as those that may be found in resource-limited settings. Unlike conventional diffusion models, the proposed residual diffusion model restores the residual between the ground truth and an initial mapping. Furthermore, in contrast to the traditional diffusion process, our approach introduces noise selectively into the foreground regions during the forward diffusion phase, specifically designed for brain images where the background outside of the brain is not essential.
- 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 authors conduct an ablation study to further dissect the impact of key model components on the image generation process. Incrementally removing elements such as mask, residual diffusion (Res), BRAN, and MAP allowed them to gauge their contribution to the model’s performance. The authors use Mean Absolute Error (MAE), Learned Perceptual Image Patch Similarity (LPIPS), Structural Similarity Index Measure (SSIM), and Peak Signal-to-Noise Ratio (PSNR) for perceptual evaluations.
- 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 paper would benefit from citing other relevant work in this area to achieve low-cost to high-cost mapping of imaging data using generative models. Some such papers include the following: https://link.springer.com/chapter/10.1007/978-3-031-18523-6_15 and https://link.springer.com/chapter/10.1007/978-3-031-06427-2_18
- 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 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?
The authors claim code is available, but I assume it will be made available upon acceptance, as it is not visible now.
- 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 paper would benefit from citing other relevant work in this area to achieve low-cost to high-cost mapping of imaging data using generative models. Some such papers include the following: https://link.springer.com/chapter/10.1007/978-3-031-18523-6_15 and https://link.springer.com/chapter/10.1007/978-3-031-06427-2_18
- 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 paper would benefit from citing other relevant work in this area to achieve low-cost to high-cost mapping of imaging data using generative models. Some such papers include the following: https://link.springer.com/chapter/10.1007/978-3-031-18523-6_15 and https://link.springer.com/chapter/10.1007/978-3-031-06427-2_18
- 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 the thoughtful feedback, and will clarify all the concerns from the reviewers in the final version. Our responses to the major points raised by the reviewers as follows.
Reviewer #1
Q1: Minor comment: in Fig. 4, it might be easier for readers to observe with colored images instead of the grey-scale visualization for perfusion maps. A1: We understand that colored visualizations can enhance readability and make it easier for readers to observe the details. We will incorporate your suggestion and update the figures accordingly in the final version.
Reviewer #3
Q1: While diffusion (resp. residual diffusion) is an established and widely concept in the general deep-learning community by now, it is still limited within the medical context. The text relies on a working knowledge of it, limiting its audience in the wider community. A1: We acknowledge that, while prevalent in deep learning, these terms might be unfamiliar to a wider medical audience. We’ll explore incorporating a concise explanation of diffusion and residual diffusion within the medical context, tailored to the specific application in our paper. This will improve accessibility for readers with varying backgrounds. Alternatively, we will reference a well-established paper titled《Diffusion models in medical imaging: A comprehensive survey》which effectively explains these concepts. This approach avoids redundancy and streamlines the paper’s focus. We’ll carefully evaluate these options to strike a balance between conciseness and clarity for a broader readership.
Q2: The work hints at clinical applicability but lacks key characteristics for this like timings and computational overhead. A2: The running time and computational efficiency of the diffusion model is often related to the setting of the time step t. The average running time of our model for various settings will be incorporated into the table of results. Additionally, the accelerated sampling methods for diffusion models are continuously improving, which makes the time and computational overhead of our algorithm controllable.
Q3: Unclear what statistics where used, i.e. are the shown results averages over the dataset, or runs etc. While common, comparing models based on single instance is not a valid approach due to the stochastic nature of training. A3: We agree relying solely on averages from a single dataset and model run can be misleading due to the inherent stochasticity of training. We will consider this suggestion to incorporate additional model runs with the reported statistics reflecting the average and standard deviation across these runs. This will provide a clearer picture of the model’s performance variability.
Reviewer #4 Q1: The paper would benefit from citing other relevant work in this area to achieve low-cost to high-cost mapping of imaging data using generative models. A1: We agree that our paper would benefit from discussing and citing other relevant work in this area. We will properly cite and discuss the references you provided to enhance our paper’s context and depth.
Meta-Review
Meta-review not available, early accepted paper.