List of Papers Browse by Subject Areas Author List
Abstract
Late gadolinium enhancement (LGE) imaging is considered the gold-standard technique for evaluating myocardial scar/fibrosis. In LGE, an in-version pulse is played before imaging to create a contrast between healthy and scarred regions. However, several factors can impact the contrast quali-ty, impacting diagnostic interpretation. Furthermore, the quantification of scar burden is highly dependent on image quality. Deep learning-based au-tomated segmentation algorithms often fail when there is no clear boundary between healthy and scarred tissue. This study sought to develop a genera-tive model for improving the contrast of healthy-scarred myocardium in LGE. We propose a localized conditional diffusion model, in which only a region-of-interest (ROI), in this case heart, is subjected to the noising pro-cess, adapting the learning process to the local nature of our proposed en-hancement. The scar-enhanced images, used as training targets, are generated via tissue-specific gamma correction. A segmentation model is trained and used to extract the heart regions. The inference speed is improved by lever-aging partial diffusion, applying noise only up to an intermediate step. Fur-thermore, utilizing the stochastic nature of diffusion models, repeated infer-ence leads to improved scar enhancement of ambiguous regions. The pro-posed algorithm was evaluated using LGE images collected in 929 patients with hypertrophic cardiomyopathy, in a multi-center, multi-vendor study. Our results show visual improvements of scar-healthy myocardium con-trast. To further demonstrate the strength of our method, we evaluate our performance against various image enhancement models where the proposed approach shows higher contrast enhancement. The code is available at: https://github.com/HMS-CardiacMR/Scar_enhancement.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2024/paper/3611_paper.pdf
SharedIt Link: https://rdcu.be/dVZev
SpringerLink (DOI): https://doi.org/10.1007/978-3-031-72378-0_29
Supplementary Material: https://papers.miccai.org/miccai-2024/supp/3611_supp.pdf
Link to the Code Repository
https://github.com/HMS-CardiacMR/Scar_enhancement
Link to the Dataset(s)
N/A
BibTex
@InProceedings{Has_Myocardial_MICCAI2024,
author = { Hasny, Marta and Demirel, Omer B. and Amyar, Amine and Faghihroohi, Shahrooz and Nezafat, Reza},
title = { { Myocardial Scar Enhancement in LGE Cardiac MRI using Localized Diffusion } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2024},
year = {2024},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15001},
month = {October},
page = {307 -- 316}
}
Reviews
Review #1
- Please describe the contribution of the paper
Deep learning-based scar segmentation models could fail on low contrast LGE cases. To solve this challenge, the authors focused on enhancing a local region containing scars by training a diffusion model. Their experiments showed that the proposed method could visually improve the contrast between scar and healthy myocardium.
- 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.
-
Brief pipeline. The authors first trained a segmentation network to identify region of interests. Based the results of segmentation, they further trained a localised diffusion model to enhance the ROIs. This two-step pipeline is brief and easy to understand.
-
Clear writing. This paper is easy to follow.
-
- 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.
-
Methodology is not sound. The authors focused on myocardial scar enhancement in LGE since low contrast could lead to failed segmentation for learning-based methods. However, in their methodology, the authors still used the segmentation results from the low contrast LGE for the myocardial scar enhancement, which is not sound. A potential issue is the proposed method cannot accurately localise myocardial scars and enhance them when the segmentation is unsuccessful on low contract LGE. That means this work still cannot solve the real-world challenge of quantifying scars on difficult LGE cases.
-
Inconsistent training and inference. During training stage, the authors simulated scar enhancement by gamma correction and fed enhanced LGE for training a diffusion model. During inference stage, the authors fed low contrast LGE to the diffusion model. Therefore, there exists domain shift between training and inference. Since the diffusion model has never seen the low contrast LGE, the robustness on the unseen cases is questionable.
-
Unconvincing experimental validation. The authors focused on improving scar contract in LGE for better scar quantification. However, there is no experimental result to show that the enhanced LGE is easier for quantifying scars. Particularly, for the LGE cases which are difficult to segment using previous scar segmentation models, the authors did not show their method could enhance these cases and achieve better segmentation. Therefore, it is hard to evaluate if the proposed method is useful in scar quantification.
-
- 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 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
-
Clarify the soundness. The proposed method depends on a segmentation model for identifying ROIs. If the segmentation model fails in low contrast LGE cases, the proposed method will fail in enhancing scar regions. The authors need to clarify why the proposed method is efficient to the difficult LGE cases.
-
Validate the effectiveness in scar quantification. The proposed method is useful if it could improve the accuracy of segmenting scars. In their future work, the authors need to provide the results for scar segmentation, especially on difficult LGE cases, to demonstrate the effectiveness of their method.
-
- 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?
The methodology of this work is not sound. The authors have not demonstrated that the proposed myocardial scar enhancement can help scar quantification.
- 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
All reviewers raised the concern on segmentation error, which has been well addressed in the rebuttal. Besides, my concern on methodology and validation have also been addressed. I recommend to accept this paper.
Review #2
- Please describe the contribution of the paper
This paper introduces a novel diffusion-based model to enhance the visualization of myocardial scars in Late Gadolinium Enhancement (LGE) Cardiac MRI. The model applies a localized conditional diffusion process focused on regions of interest (ROI), specifically the heart, to improve the contrast between healthy and scarred myocardial tissues. The proposed method leverages gamma correction for target data preparation and partial diffusion to expedite inference, making the process computationally efficient and potentially more applicable in clinical settings.
- 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 application of localized diffusion to enhance myocardial scar visualization in LGE MRI is novel and shows a significant improvement in image quality. This localized approach ensures that enhancements are concentrated where they are most needed, preserving the rest of the image’s integrity. 2.The approach is validated on a substantial dataset from a multi-center, multi-vendor study, which supports the generalizability of the findings. The extensive performance comparison with other image enhancement methods underscores the superiority of the proposed model in enhancing myocardial scar visibility.
- 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 method is innovative, the complexity of the implementation might limit its accessibility. More detailed discussions or simplifications on the integration into clinical workflows could enhance the paper’s practical value.
- The effectiveness of the model relies on heart segmentation. Errors in segmentation could lead to less effective enhancements. Exploring the robustness of the model to segmentation errors or providing solutions to mitigate such issues could strengthen the paper.
- The paper is well-structured and written, but some sections are highly technical. Simplifying these parts or providing more explanatory figures could help in reaching a broader audience, including clinical practitioners.
- 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
- More information on how the enhanced images affect diagnostic outcomes would be highly beneficial. If possible, include preliminary clinical feedback or simulate a study on diagnostic accuracy with and without the enhanced imaging.
- Propose a framework for a long-term clinical study that could evaluate the impact of these enhanced images on patient outcomes, particularly focusing on the management and treatment decisions for conditions like hypertrophic cardiomyopathy.
- Discuss the potential of using fully automated segmentation methods to reduce human input and increase reproducibility. If manual segmentation is required, suggest strategies or tools that could help streamline this process.
- 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 experimental results.
- 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
This study introduces a localised conditional diffusion model to enhance the contrast between scarred and healthy myocardial tissue in Late Gadolinium Enhancement (LGE) imaging. The model selectively applies noise to defined regions of interest and uses partial diffusion to improve inference speed while maintaining image quality. The effectiveness of this approach is validated through comparative analysis with other image enhancement techniques using a dataset from a multi-centre, multi-vendor study involving 929 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 introduction of a localised conditional diffusion model represents a novel methodological contribution to enhancing contrast in LGE images.
- The paper validates a methodological innovation through the use of partial diffusion, which limits the diffusion process to an intermediate step rather than completing a full cycle, useful in medical imaging.
- The paper stands out for its rigorous evaluation framework that benchmarks the proposed model against multiple image enhancement techniques, including CNNs and GAN-based models.
- 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 efficacy of the localised diffusion model is highly dependent on accurate segmentation of the left ventricle. Incorrect segmentation can lead to misapplied enhancements, potentially reducing the clinical utility of the processed images.
- The paper claims significant improvements in image quality and diagnostic utility, yet it does not provide comprehensive statistical tests to support these claims.
- While the paper demonstrates the application of the localised diffusion model to LGE images from patients with hypertrophic cardiomyopathy, it does not discuss or validate the model’s effectiveness across different cardiac conditions or other imaging modalities.
- 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?
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
-
While there are existing methodologies for improving scar-healthy myocardium contrast post-acquisition, such as leveraging T1 maps and cine images to derive LGEs with enhanced image quality (https://doi.org/10.1161/CIRCULATIONAHA.122.060137), this study focuses specifically on enhancement techniques applied directly to post-acquisition LGE images.
-
The variability in intensity distributions among LGE images, depending on the imaging sequence employed, raises questions about the type of LGE images used in this study. It is necessary to specify whether the study used magnitude-based images or PSIR-based images and to discuss the applicability of the proposed method to out-of-distribution data.
-
The effectiveness of the proposed localised diffusion framework relies on accurate segmentation of the left ventricle. The impact of incorrect segmentation predictions on the enhancement process should be discussed, including the frequency and implications of such errors.
-
The pre-processing step requires images centred on the left ventricle. In practice, the deployment of this framework to new, unseen cases would necessitate reliable automated methods for centring images, raising concerns about potential errors propagated from predicted left ventricular masks.
-
The study focuses on the localised region of interest; however, the impact on the rest of the image remains unclear. It is important to discuss whether only the specific ROI is enhanced and how this affects the overall appearance and intensity alignment of the full image.
-
The selection of lambda (λ) for partial diffusion at a value of 250 should be clarified, particularly whether this value was empirically determined across the entire training set.
-
The methodology for selecting lambda (λ) for the unconditional DDPM used in comparative analyses needs to be detailed to understand the basis of this choice.
-
Discrepancies between the references cited for comparison methods and those presented in Table 1 should be addressed to ensure accurate attribution and verification.
-
The claim of significant improvements in image enhancement lacks statistical validation. It is essential to describe the statistical tests used to support these claims.
-
The advantage of shortened inference times through partial diffusion is noted, but specific data on these times are missing. Detailed reporting of inference times and the number of iterations required to process one image would provide a clearer understanding of the method’s efficiency.
-
Validating the enhanced scar visibility using different datasets and assessing the ease of scar segmentation in enhanced images would substantiate the framework’s utility and effectiveness.
-
For ease of verification and access, it is advisable to order the references as they appear in the text rather than alphabetically, ensuring consistency and ease of reference throughout the paper.
-
- 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 recommendation to accept this paper is primarily justified by its introduction of a novel localised conditional diffusion model tailored for enhancing the contrast in LGE images. Additionally, the validation of partial diffusion to expedite the inference process addresses a common practical limitation in clinical settings, providing a pragmatic improvement over existing methods. The evaluation of the model against established baselines using a large, multi-centre dataset further supports the paper’s conclusions and demonstrates the clinical applicability of the approach. Despite some limitations, such as the dependency on accurate segmentation and the lack of extensive statistical validation, the strengths of the methodological innovation and its effective demonstration of utility in a clinical context are compelling reasons for acceptance.
- 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 other reviewers added very valuable comments. The authors have addressed all major comments successfully.
Author Feedback
We thank all the reviewers for their supportive feedback, insightful, and valuable comments. We are thrilled that the reviewers find our proposed methodology novel (R4, R7) and easy to understand (R6). Moreover, we are grateful for the acknowledgment of our rigorous (R7) and extensive (R4) performance comparison in a multi-center study and the methodological innovation of partial diffusion (R7). Impact of Heart Segmentation (R4, R7): We acknowledge the importance of accurate heart segmentation, but our model performance does not entirely depend on it. Segmenting the heart in LGE imaging is easier than scar segmentation, and commercial software exists for this task. Mitigation strategies, like dilation, can slightly enlarge the ROI at a low computational cost to ensure heart inclusion. Clinicians can use manual landmark or “rough” segmentation as a starting point and easily verify and correct the segmentation if necessary. We will address this point in the discussion section. Clinical Impact (R4): Our model could substantially improve LGE image interpretation and quantification. Despite advances in LGE sequences, ambiguity remains in 20-25% of cases regarding the presence, absence, and extent of scars, especially in patients with nonischemic cardiomyopathy. Contrast enhancement could guide readers to high-probability scar areas, enhancing CMR/LGE’s global reach by simplifying image interpretation in centers without CMR experience. Our method also contribute to the standardization and automation of scar burden in LGE- an unmet clinical need. For example, in HCM patients, we use a 15% threshold of scar burden to assess arrhythmia risk, yet scar quantification remains a well-known challenge. We will add insights into the clinical utility of enhanced images and their impact on image interpretation and patient management in the introduction and discussion. Implementation Complexity (R4): Our model can be easily integrated into clinical workflow, either on the MRI system as part of reconstruction/image enhancement or on image PACS for a vendor-neutral approach. We have extensive experience in this area and plan to integrate the model into our scanner upon completion of our technical development. Similar methods have been commercialized by Subtle Medical for neuro MRI/nuclear imaging. The model was trained on a single GPU and does not require extensive computational resources. We are making the code and the model publicly available for easy integration by other teams. As suggested, we will discuss the potential of fully automated segmentation methods and the framework’s utility in long-term clinical studies to evaluate the impact of the framework on patient outcomes. Scar Quantification (R6): We acknowledge that we did not fully evaluate the model for scar quantification, as this paper is not about scar segmentation- though scar segmentation is a byproduct of improving contrast enhancement. We believe there may have been a misunderstanding of the framework. Scar segmentation was used solely to generate ground truth and apply gamma correction for the training set, performed manually. The model’s input is the heart, so no scar segmentation is required. We will further clarify this important issue in our manuscript. Inconsistency between training and Inference (R6): We follow the typical process for training image-to-image diffusion models, where the denoising UNet is trained using target data to denoise images in the target domain. The model is conditioned on acquired LGE images during training and inference, enhancing its ability to map images from the original to the enhanced domain. The framework was evaluated using a large multicenter, multivendor dataset, as acknowledged by R4, with results reported on unseen cases. Statistical Analysis (R7): We will include additional statistical tests and adjust the manuscript accordingly. We thank the reviewer for pointing out this important shortcoming of our manuscript.
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’
All reviewers are happy to accept this work after the rebuttal.
- 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).
All reviewers are happy to accept this work after the rebuttal.
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