List of Papers Browse by Subject Areas Author List
Abstract
The advancement of electron microscopy (EM) imaging technology has expanded its applications in life science research, making the automation of EM image analysis a key focus in biomedical imaging. As a core task in EM image analysis, semantic segmentation has garnered significant attention, and convolutional neural networks (CNNs) have been extensively studied, currently emerging as the mainstream method. However, existing methods still face several unresolved challenges. One issue arises from the convolution process, which makes it difficult to efficiently balance global and local information, thus limiting further improvements in segmentation accuracy. Another issue stems from the nature of CNNs, which aim to establish an optimal mapping between images and labels, achieving high accuracy in in-domain data segmentation but at the cost of a noticeable performance drop on out-of-domain data. In this paper, we explore the potential of diffusion probabilistic models (DPMs), known for their exceptional image modeling capabilities, to address these challenges. Specifically, we introduce a diffusion probabilistic model for the semantic segmentation of EM images, which we call EM-Cold-SegDiffusion (ECSD). We adopt a cold or deterministic diffusion framework to achieve higher inference efficiency and a more deterministic segmentation process. Additionally, by introducing an edge-sensitive loss function, we significantly enhance both training efficiency and model performance. Experimental results on common EM segmentation tasks demonstrate that ECSD outperforms mainstream models, offering a promising and superior solution for EM segmentation.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/0977_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{QiMug_ABoundaryaware_MICCAI2025,
author = { Qi, Muge and Shi, Ruohua and Cai, Yu and He, Liuyuan and Wang, Wenyao and Ma, Lei},
title = { { A Boundary-aware Cold-Diffusion Model for Electron Microscopy Segmentation } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15975},
month = {September},
page = {12 -- 22}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper applies Cold SegDiffusion [21] to the EM segmentation task. It proposes a Time-Adaptive Boundary Attention, based on Dermosegdiff [5], as a weight of the loss function.
- 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 paper proposes a Time-Adaptive Boundary Attention weight with a time-decay factor to dynamically adjust for each time step ( t ). It explains how to create the attention weight map for each EM image through a sequential process: blur the image, run edge detection, calculate the distance map, and invert the distance map.
- 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.
1) From the beginning, the paper presents two key issues: “local and global information” and “data generalization (or OOD)”. The former issue seems to be addressed by using DPMs and their attention weight map. However, the latter issue remains unresolved due to the lack of further testing on an OOD dataset. Furthermore, although the paper is aware of the Segment Anything Model (SAM) [11] and its variants, it does not include any experiments involving them, which raises a concern, as comparing with SAM would provide more meaningful insights given its status as a well-known foundation segmentation model. Additionally, if the paper mentions transformer-based models, it should include at least one for comparison to support its claims.
2) Concern about the clarity of EM-Cold-Seg-Diffusion: The paper explains that in the diffusion process, x_0 is the “original medical image (input image)”, while Figure 2 shows that x_0 is the segmentation mask. Similarly, x_T is described as the masked image, but it appears as the EM image in Figure 2. Furthermore, since this paper is an application of Cold-Seg-Diffusion, Figure 2 should at least indicate the locations of the Contrast Enhancement Module (CEM), Inverse Fast Fourier Transform (IFFT), the Channel Attention Mechanism (CAM), and Spatial Attention Mechanism (SAM). It should also distinguish between the segmentation encoder and the conditional encoder, as well as their respective inputs.
3) In Section 2.2, the paper appears to use functions from OpenCV but does not mention or explain this. It directly shows “cv2.Canny” and “cv2.distanceTransform” without providing any clarification or context.
4) The arrows in Figures 4 and 5 are inconsistent and not properly aligned across methods. For example, in the last row of Figure 4, the red arrow is small in “GobleNet” and “Cold SegDiffusion”, but in “Our”, it is larger and points in a different direction. Furthermore, Figures 4 and 5 appear to show only cherry-picked examples and do not include failure cases. For instance, in Figure 4, the second row of “Our” shows a false segmentation in the top-left corner that does not appear in the ground truth (GT) mask. In the third row of Figure 4, the left edge of the GT mask includes a segment that is missing in “Our”. Similarly, in Figure 5, the first row shows a blue segment on the left edge in “Our”, while the GT mask is yellow in that region.
- Please rate the clarity and organization of this paper
Poor
- 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.
- 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
1/ The paper contains multiple typos: “The The”, “STo”, “[20,16,20]”, “DISTL2”, “4 × 10”, “1 × 10”, “the++ dataset”, Fig1: “DNcoder”. 2/ In the Introduction, the paper explains the abbreviation of DPMs but does not explain ECSD again, although both are introduced in the abstract. 3/ The paper lacks experimental details despite having available space. It does not specify the cropping size or training size, and the time step “T” is not clearly defined (T is set to 50 in Cold SegDiffusion [21], but it should still be mentioned explicitly, as it is a hyperparameter).
4/ Figures 4 and 5 should be revised to improve their quality. There is one white gap line in Figure 4 but two lines in Figure 5, which is inconsistent. Additionally, it is unclear why there are blue boundaries on the EM image in the first row.
5/ The paper does not explain how the value “1.1415” is calculated. This should be clarified for reproducibility.
6/ Fig 3 - 1 show the Blur image but it is not explained in the paper. It might be included in the opencv
- 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?
Based on my above comments, despite the fact that this paper is an application of Cold SegDiffusion [21] to the EM segmentation task and proposes a novel attention weight map to improve performance, the paper has several issues in terms of writing, addressing the mentioned problems, and providing necessary clarifications.
- Reviewer confidence
Very confident (4)
- [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.
Thank you to the authors for taking the time to address all my questions and revise the paper. As Reviewer 1 and I mentioned earlier, this work combines concepts from Cold SegDiffusion [21] and Dermosegdiff [5], but it makes a meaningful contribution by exploring their application to EM segmentation. I hope the authors will consider open-sourcing their code so that others can experiment with their method.
Review #2
- Please describe the contribution of the paper
The paper explores the use of a diffusion probabilistic model DPM applied for the first time for the semantic segmentation of EM images. They combine the DPM with an edge-sensitive loss function and demonstrate its improvements for both training efficiency and segmentation performance. The paper validates the approach comparing the segmentation performance across two datasets and tasks (mitochondria segmentation and multi-class organelle segmentation) outperforming a state-of-the-art EM segmentation model from the literature and showing the performance gain when using the edge sensitive loss.
- 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.
Demonstrated the successful application of diffusion probabilistic models to the challenging problem of EM semantic segmentation.
- 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.
Limited novelty as the introduced network is the same as the previously proposed Cold SegDiffusion (Yan et al.) with a modified loss that is proposed in DermoSegDiff (Bozorgpour et al.).
- 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.
- 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?
Strong results were demonstrated on two different datasets and tasks. However, as the paper appears to directly combine two existing methods, the novelty is somewhat limited. Still, I’m impressed by the results and design decisions of the authors.
- 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.
The rebuttal has further improved the paper, putting it in better context with existing work and showing the advantages of the method.
Review #3
- Please describe the contribution of the paper
The paper discusses the application of diffusion models in the context of Electron Microscopy(EM) segmentation. The authors built an EM segmentation algorithm derived from Cold Diffusion and ColdSegDiffusion and propose an edge-sensitive regularization technique that significantly improves the model’s performance
- 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 authors identified the potential of diffusion models in medical image segmentation and provides valuable addition to the knowledge of the medical imaging field as it evolves beyond the mainstream paradigm of UNet
- The core contribution of the paper is the boundary-aware loss regularizer, in which the authors used cv2 edge detection to provide an unsupervised heat map of inverse distance-to-nearest-edge as the regularizer. Such that the loss contribution around edge pixels are more salient. Another interesting addition is the time decay factor such that as diffusion approaches the end, the loss terms relaxes attention to non-edge pixels. I find the intuition behind this change quite insightful and the metric gains quite substantial.
- 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.
- In terms of novelty, because the major backbone of the segmentation algorithm is CodeSegDiffusion, the overall contribution is still largely incremental. Some recent publications may also be exploring the time dependent adjustment strategy in diffusion model but they are relatively concurrent and quite different from what this paper is trying to address. https://onlinelibrary.wiley.com/doi/full/10.1002/ima.70067
- The authors only reported using cv2 for boundary detection, while straightforward and lightweight, it may not be optimal accuracy-wise, it would be interesting to see what happens when a more powerful boundary detector is deployed(e.g. a supervise-trained one would be interesting to compare against)
- 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.
- 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.
(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?
In EM segmentation, one major challenge has been the long-tail hard boundary examples, where borders tend to become blurry or corrupted due to imaging limitations. False negative detections of boundaries can incur outsized harm to downstream tasks like 3D object segmentation. The boundary conditioned regularization addresses this problem and demonstrates convincing improvements on segmentation quality.The boundary conditioned regularization addresses this problem and demonstrates convincing improvements on segmentation quality.
- Reviewer confidence
Very confident (4)
- [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.
The authors will add more citations to relevant work
Author Feedback
We sincerely thank the AC for overseeing the review and providing the rebuttal opportunity. We appreciate R1 and R3 for acknowledging the practical value of our work. We especially appreciate R3’s thoughtful understanding of our intentions, as well as the constructive suggestions. We also thank R2 for the valuable comments on experimental design and for carefully pointing out the typos in the text and figures. #R1:Q1: We believe our contributions go beyond a simple combination of existing methods, for the following reasons: Notable performance improvement: To our knowledge, this is the first attempt to apply diffusion models to EM segmentation, and the performance gains are substantial. Tailored loss strategy: The time-dependent adjustment of loss functions shows great potential in diffusion models. We have also explored alternatives, but the current setup performed best. Additionally, considering EM image characteristics, we incorporated a Gaussian blur step to suppress non-membrane textures before edge detection, which proved effective. Domain significance: We believe MICCAI values both algorithmic innovation and domain relevance. Our work addresses the unique challenges of EM segmentation and provides a validated solution for the field. #R2:Q1:
We agree with R2’s suggestion of experiments on OOD data. We have conducted cross-validation between MitoEM-R and MitoEM-H. Our method achieved Dice scores of 75.23% (R→H) and 80.35% (H→R), outperforming the NoAdapt (Wu et al., 2021 MACCAI) by 18.43% and 3.7%, respectively. The comparison with SAM and Transformer-based models has been conducted in the GobletNet paper and is now included in our Table 1. On BetaSeg, our method achieves a Dice score of 83.09%, outperforming SwinUnet (47.37%) and SAM (48.99%). Q2: We have corrected in Section 2.2: “Diffusion Process: EM mask images (input label) are progressively degraded to EM original images, …” And we have added the CEM, IFFT, CAM, and SAM modules in Figure 2. Q3: We have corrected the equations: edge = Canny(x_blur, \tau_low, \tau_high); distance = DT(edge). And added the clarification: “Canny is a common edge detection method used to locate edge pixels in images[2], …”, “Then, the Euclidean distance transformation algorithm[3] is used to calculate the distance of each pixel to the nearest non-zero pixel …” ([2][3] are added refs.) In sec.3.2, we added the context: “The Canny and DT methods are implemented by the OpenCV library.” Q4: We apologize for inconveniences caused by the figure issues; however, these do not affect the conclusions. First, the figures were randomly selected and are representative. We have marked our failure cases in the revised version. Additionally, we will standardize figure annotations and legends (Blue, Red, Yellow, and Green indicate errors from all models, unique to GobletNet, unique to Cold-Seg-Diffusion, and unique to ESCD. Optional: We have revised the paper accordingly:
- Check and correct all the typos.
- Modify Introduction: “… our EM-Cold-SegDiffusion (ECSD) based…”
- Add experimental details in Sec. 3.2, including T, batch size, crop size.
- Reorganize the figures.
- Add the explanation in Sec. 2.2: “1.1415 is √2 inherited from Dermosegdiff”
- Include a description of the blurring operation. #R3:Q1: We sincerely thank you for recognizing our work. We fully agree that time-dependent loss adjustment is key to further unlocking the diffusion model’s potential. Our paper uses a time decay factor to emphasize global structures early and refine local textures later. Similarly, the DPCT’s TDC module mentioned in your reference offers valuable insights. We plan to try it in our model and will continue to follow advancements in this area. Q2: We appreciate your suggestion. Although OpenCV offers an efficient solution, we also acknowledge its limitations in accuracy. Investigating more advanced boundary detection methods has become an important direction in our continued work.
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”.
N/A
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- 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