List of Papers Browse by Subject Areas Author List
Abstract
In recent years, accelerated MRI reconstruction based on deep learning has led to significant improvements in image quality with impressive results for high acceleration factors. However, from a clinical perspective image quality is only secondary; much more important is that all clinically relevant information is preserved in the reconstruction from heavily undersampled data. In this paper, we show that existing techniques, even when considering resampling for diffusion-based reconstruction, can fail to reconstruct small and rare pathologies, thus leading to potentially wrong diagnosis decisions (false negatives). To uncover the potentially missing clinical information we propose ``Semantically Diverse Reconstructions’’ (\SDR), a method which, given an original reconstruction, generates novel reconstructions with enhanced semantic variability while all of them are fully consistent with the measured data.
To evaluate \SDR automatically we train an object detector on the fastMRI+ dataset. We show that \SDR significantly reduces the chance of false-negative diagnoses (higher recall) and improves mean average precision compared to the original reconstructions. The code is available on \href{https://github.com/NikolasMorshuis/SDR}{https://github.com/NikolasMorshuis/SDR}.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/1731_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/NikolasMorshuis/SDR
Link to the Dataset(s)
N/A
BibTex
@InProceedings{MorJan_Mind_MICCAI2025,
author = { Morshuis, Jan Nikolas and Schlarmann, Christian and Küstner, Thomas and Baumgartner, Christian F. and Hein, Matthias},
title = { { Mind the Detail: Uncovering Clinically Relevant Image Details in Accelerated MRI with Semantically Diverse Reconstructions } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15961},
month = {September},
page = {354 -- 364}
}
Reviews
Review #1
- Please describe the contribution of the paper
Existing techniques may fail to reconstruct small pathologies when MR data is highly undersampled. This study proposes a method called Semantically Diverse Reconstruction, which generates images with enhanced semantic variability. The resulting images maintain consistency with the acquired data while potentially revealing clinically relevant pathologies.
- Please list the major strengths of the paper: you should highlight 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 study integrates pathology detection and enhancement into MR reconstruction. The idea is conceptually interesting.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
My major concern is that the proposed method may have limited clinical applicability.
-
The method aims to generate a set of reconstructions that are semantically diverse in regions where pathologies are likely to occur. However, I am uncertain about the necessity of this diversity. If the pathology is consistent, a single accurate image should suffice for diagnosis. If there are differences in the reconstructed pathologies, it raises the question of how to determine which one reflects the true condition.
-
In principle, the reconstructed image from fast MR imaging techniques should be as close as possible to the ground truth. Otherwise, the acceleration rate should be reduced to maintain consistency. Artificial features introduced during reconstruction may interfere with diagnosis. For example, in Figure 1 (bottom right), the two example images show substantial differences. Although one appears closer to the ground truth, many details are missing, suggesting limited practical utility.
-
Regarding the box proposal selection, if the initial reconstruction misses certain pathologies, it is unclear whether the detection method can still accurately localize the affected regions. If it can, does that imply the image already contains sufficient information for diagnosis?
-
- 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.
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html
- Please refer to the weaknesses mentioned above.
- The visual differences in Figure 5 are minimal; it is recommended that professional radiologists assess the diagnostic relevance of these variations.
- 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.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
My recommendation is based on concerns regarding the clinical applicability of the proposed method. While the idea of generating semantically diverse reconstructions is interesting, its necessity and reliability in clinical scenarios remain unclear. The potential introduction of artificial features, lack of clear ground truth when multiple reconstructions differ, and uncertainty in pathology localization raise significant doubts about its practical utility.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
In this paper, a reconstruction approach for accelerated MRI is proposed that generates multiple, diverse image candidates, all consistent with the measured data. It uses a vision-transformer based notion of image distances, and tries to generate samples far apart. The method is evaluated on the fastMRI+ dataset, showing higher recall and precision than conventional reconstruction methods.
- Please list the major strengths of the paper: you should highlight 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.
One strength of this paper is the interesting concept, with the notion of generating a set of “semantically diverse” images, with a distance metric based on vision-transformer features. This gives a broader set of possible reconstructions, than multiple samples from a diffusion-based generative model, or dfferent initializations of a CNN reconstruction model.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
The weakness of this paper is that it is unclear what an end user (e.g. radiologist) would be expected to do with the set of images. If the proposed method generates 2 outputs, one with a lesion and one without, how is this meant to guide decision-making? How would the radiologist know which image is correct? There is no clear way to utilize these outputs.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html
N/A
- 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.
(4) Weak Accept — could be accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper shows interesting results, and presents a novel and interesting idea (generating a set of semantically diverse, data-consistent reconstructions). However, the ultimate utility of the approach is unclear.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
Authors propose “Semantically Diverse Reconstructions” (SDR), a method which, given an initial reconstruction (using any method that provides aliasing-resolved reconstruction from undersampled data), can generate a set of reconstructions that are semantically diverse in automatically or manually selected box regions where pathologies are likely to occur and at the same time maintain consistency with the acquired undersampled k-space. The diverse reconstructions are created using gradients from feature encoder of a ViTDet detection network, which is adversarially robust. Compared to a single reconstruction from conventional DL recon (which can be blurring and miss small pathologies) and random samples from diffusion models (which may tend to generate healthy tissue due to larger portion of healthy data in the training), the diverse reconstructions have better opportunity to show potential pathologies consistent with the acquired data. The approach was evaluated in a pathology detection task in fastMRI+ knee data. It outperformed the initial reconstruction and random samples in pathology recall, while mean average precision is not affected.
- Please list the major strengths of the paper: you should highlight 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.
- Clinical importance: This paper is tackling an important problem in medical imaging reconstruction, i.e., preservation of clinically relevant information. Currently most existing reconstruction work has been focusing on overall image quality (e.g., PSNR, SSIM) in accelerated MRI. It is valuable to pay attention to presentation of small and rare pathologies.
- Novelty: It is an interesting idea to create semantically diverse reconstructions from the same undersampled k-space data. In accelerated MRI reconstruction with multiple possible solutions, conventional DL methods tend to generate averaged results whereas more recently diffusion models provide random samples from the learned distribution. Although people have been using the multiple samples from diffusion model to estimate uncertainty, few works have tried to intentionally create different samples with improved clinical value. This work shows a new possible direction.
- Flexibility: The proposed approach can work with different initial reconstructions. Authors tested the approach with three different initial reconstruction methods and showed consistent improvement in pathology detection. Additionally, authors provided two options to create the boxes, either automatically or manually.
- Fast speed: The approach takes 3 seconds per image in generation, which is faster than diffusion samplings and looks clinically feasible.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
- Reliance on box generation: The method requires defining candidate box regions as the first step for diverse reconstructions on these regions. These boxes need to be defined on the initial reconstruction, where the pathologies are very likely to be missing. This can be a challenging task. In the manual box generation experiment, boxes were generated by perturbing the ground truth, which is a relatively ideal case. In practice, it may require experience of radiologists to draw boxes on regions where they cannot see the pathology but still find them suspicious. The quality of this step may affect the final results.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html
Thanks for the authors’ efforts and it is a nice work. Just some thoughts to share (not requests): It is interesting to see a ViTDet detection network is used to obtain gradient, which is further used to generate new reconstructions semantically different than others. Would the new features created in this way be medically realistic? I notice that ViTDet was trained on ImageNet. There may be a mismatch in semantic information between natural images and specific medical images. Would finetuning on fastMRI be helpful to improve the reality? While the gradient is based on box-features, will images outside of the boxes be affected? Is box selection a necessary step in this workflow and will the method work on whole image? Some evaluation from radiologists will be helpful in the future to have a better understanding about the generated image quality and how they will utilize these different reconstructions for diagnosis.
- 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.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The ideas of paying attention to preservation of clinically relevant information and creating semantically diverse reconstructions that may help detect pathology in undersampled MRI reconstruction are something that researchers in this field can learn from. The evaluation results look convincing, and the paper presentation justifies itself.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
My initial review was more from the perspective of technical development, and I found the idea of generating diverse samples interesting, especially when diffusion samplings become popular. It provides a different viewpoint in terms of generating multiple samples and uncertainty estimation. After reading other reviewers’ reviews, I agree with them that the practical utility of this method is unclear. The paper uses automatic detection methods to show that the method helps reduce false negative of the detection model. However, in clinical practice it remains unclear how radiologists should make diagnosis given the diverse reconstruction results. I agree with Reviewer #1 that in this case more measurements are needed to confirm the results. The authors’ response mentioned that ‘the method is the first step towards an automatic method that only stops the acquisition process when there is no more semantic uncertainty in the reconstructed image’, which is an interesting direction but not yet included in this work. The response mentioned that ‘our method does not work well without box-selection when only using the ViT-backbone, as image-wide patterns can occur’, which makes me a little concerned about the reality and robustness of the generated reconstructions. Generally, I still suggest a weak acceptance mainly because this research and the related discussion may be interesting to other researchers in the community.
Author Feedback
We thank all reviewers for their valuable feedback. We are pleased that the reviewers found the proposed idea of generating semantically diverse reconstructions interesting [R1,R2,R3], novel [R2,R3], and supported by convincing evaluation [R3]. Moreover, we are glad that [R3] believes that researchers in this field can learn from our approach. The reviewers also have raised concerns that need further clarification that we like to provide below.
[R1,R2] Practical utility: What benefit does our method provide in practice? [R1] Reconstruction should be close to ground-truth, otherwise reduce acceleration Our method has two main benefits: It reduces the chance for false negative diagnosis by finding alternative data-consistent reconstructions that differ semantically and might show pathologies that would otherwise be missed (see higher recall in Figure 4). These semantically different reconstructions also make the inherent uncertainty of the reconstruction transparent (see Section 6). The clinician can thus get an idea of the reliability of the reconstruction and, if necessary, initiate additional measurements based on semantically different images. We also see no contradiction with the reasoning of [R1], as we believe our method is the first step towards an automatic method that only stops the acquisition process when there is no more semantic uncertainty in the reconstructed image (this can happen while the person is in the scanner). We will address this point by adding a section ``Future Work’’.
[R1] Differences between reconstructions, missing details wrt. ground truth Our method aims to visualize potential semantic differences between reconstructions to show the inherent ambiguity. Therefore, semantic differences between data-consistent reconstructions is a feature. As these inverse problems have infinitely many data-consistent solutions (see Section 3.1), small differences to the ground-truth can occur in MR-reconstruction, especially for the high acceleration in Figure 1 (12x).
[R1,R3] Box-selection: Can boxes be localized automatically and manually? Does a single image contain enough information for diagnosis if boxes can be localized? Is box-selection necessary? Automatic: The region proposal network (RPN) is part of the ViTDet and proposes 1000 boxes for every image even on healthy tissue (see Figure 2). Additionally, the ViTDet adjusts the size and position of the proposed boxes during prediction, therefore the proposed boxes do not need to be perfectly accurate. Manual: A doctor might suspect a pathology in a specific area either based on the initial reconstruction or based on the patient’s indication. We will clarify the second use case in the paper. A single reconstruction might have enough information for diagnosis, but without information about its inherent ambiguity it is unknown if a confident diagnosis can be made. We have observed that our method does not work well without box-selection when only using the ViT-backbone, as image-wide patterns can occur.
[R3] Why trained on ImageNet? Newly created features medically realistic? We first train the ViTDet on fastMRI+ data and then fine-tune the ViT-backbone on ImageNet in order to make it adversarially robust. If the fine-tuning step was also performed on fastMRI+ instead of ImageNet we observed the detection scores for small pathologies to decline. Robust models are known to have semantically more meaningful gradients (see Section 3.3) and we indeed find that the reconstructions appear realistic.
[R1] Radiological assessment We agree that radiological assessment would be helpful and plan this for future work. As mentioned in Section 4, it is difficult to collect evaluations from clinicians and instead we use an object detector as a proxy for human annotators. We like to stress that the detection model used for evaluation (Faster-RCNN) is different from the one used for diverse reconstruction (ViTDet) yet it also finds the generated semantic changes.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.
The reviewers broadly agree that the idea of semantically diverse reconstructions for accelerated MRI is original and addresses a clinically relevant limitation of conventional reconstruction methods.
However, several important concerns were raised that warrant clarification. The primary issue is the clinical interpretability of multiple reconstructions—particularly when they differ in the presence of a lesion. Reviewers question how such outputs are intended to support diagnostic decisions, and what specific advantages this diversity provides in practice.
- 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
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
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