List of Papers Browse by Subject Areas Author List
Abstract
Organelle segmentation is crucial for understanding the morphology of biological structures. Existing unsupervised methods leverage powerful feature extractors and clustering techniques to uncover organelle structures from volumetric electron microscopy images. However, these methods often struggle with noisy microscopy images and the computational complexity of numerical clustering. In this paper, we propose CS2C, a novel collaborative spatial and spectral deep neural clustering framework, for multi-class organelle segmentation. The pillar of our approach is combining unsupervised deep spectral clustering and spatial clustering, which enhances a harmony of learned cluster assignments under the spatial and spectral superpixel-wise representation. Specifically, we adopt a masked autoencoder-based feature extractor to obtain powerful superpixel features, where spatial clustering is performed directly on these features. Beyond that, spectral clustering is applied in the spectral domain, naturally alleviating high-frequency perturbations in the image features. The entire framework is trained end-to-end using a combination of clustering loss and consistency regularization between spatial and spectral clustering. Extensive experiments demonstrate that our method outperforms state-of-the-art unsupervised methods on known benchmarks. Code is available at: https://github.com/JimaoJIANG/CS2C.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/2448_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/JimaoJIANG/CS2C
Link to the Dataset(s)
BibTex
@InProceedings{JiaJim_CS²C_MICCAI2025,
author = { Jiang, Jimao and Pei, Yuru},
title = { { CS²C: Collaborative Spatial and Spectral Neural Clustering for Organelle Segmentation from Volumetric Electron Microscopy } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15962},
month = {September},
page = {131 -- 141}
}
Reviews
Review #1
- Please describe the contribution of the paper
1 Integrates spatial clustering (local feature similarity) and spectral clustering (global graph structure) to address noise sensitivity and cross-image consistency. 2 Superpixel decomposition reduces computational complexity compared to pixel-level methods. 3 State-of-the-art results on BetaSeg dataset, outperforming prior unsupervised methods (e.g., MAESTER, DeepCUT).
- 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.
1 Uses GCNs to approximate spectral bases, mitigating high-frequency noise. 2 Aligns spatial and spectral clustering assignments via KL divergence, enhancing robustness.
- 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 Evaluated only on BetaSeg (mouse β-cell VEM images). Generalization to other organelles/species is untested. 2 Cluster number k significantly impacts performance. No automated method for selecting k is proposed. 3 Limited analysis of how spectral embedding captures organelle structures (e.g., visualizing spectral bases). 4 Most baselines (e.g., MAESTER) use pixel-level features, while CS2C uses superpixels. A fairer comparison would test baselines with superpixels.
- 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 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.
(3) Weak Reject — could be rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The proposed method is not highly novel, and its results show only marginal improvements over the baseline.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
The authors have not provided a reasonable parameter selection strategy, for example regarding the choice of cluster number 𝑘 k and other key settings. The claim that “the neural spectral embedding loss, clustering loss, and spatial-spectral consistency regularisation hyperparameters are all empirically set to 1” is not appropriate. Furthermore, only a limited number of clustering methods have been used as baselines; the rationale for selecting these particular methods as baselines has not been clearly explained. Given the wide range of recent advances in clustering, it is concerning that the most up-to-date methods have not been comprehensively compared. Therefore, I recommend that the paper be rejected.
Review #2
- Please describe the contribution of the paper
The proposed method introduces an unsupervised two-branch clustering framework that combines spatial and spectral clustering with consistency regularization for organelle segmentation from volumetric EM images. The spatial clustering branch performs clustering directly on semantic superpixel features extracted by a masked autoencoder model, while the spectral clustering branch conducts neural spectral embedding and clustering in a low-dimensional space spanned by the approximated spectral basis of the graph Laplacian matrix. The two clustering losses are combined in a weighted manner, requiring the tuning of two hyperparamters.The proposed method is compared to four recent approaches as well as k-means in a 3-cluster segmentation on a dataset consisting of four volumetric images of which three are used for training, one for testing. Segmentation is evaluated by Dice coefficient and IoU.The experiments show that this approach outperforms state-of-the-art techniques.
- 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.
- Very important and challenging application
- The paper is well and clearly written
- Code is available
- An ablation study is performed wrt to the two clustering branches, demonstrating the positive contribution of both approaches
- A sensitivity analysis regarding the number of clusters k is performed
- The performance gain on the chosen dataset are 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.
- The main weaknes of the paper is that only one dataset is used although the approach is general and not data-specific. Conclusions drawn from this study would be much stronger if another dataset was used to support the results
- Two additional hyperparameters are introduced but sensitivity is not discussed, they are set to 1
- Further evaluation measures assessing how many objects were missed, in particular in the case of small objects like the granules and mitochondria would be interesting to report
- 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.
- 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
Minor Comments:
- Fig. 1 caption: “adopts” -> “adopt”
- 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?
It is a well written paper with good results of the proposed method and sufficient exploration of the proposed method, but with the major weakness of testing only on a single dataset that is also very small w.r.t. independent testing data (three volumes to train, one to test).
- 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 reviewer would like to thank the authors for the additional clarifications. In summary, the manuscript provides interesting methodology and results, but would have benefitted from more evaluation, both in measuring how many objects were missed with the proposed method (as objects are of varying size) and testing the method on more than one datasets to support the general claim. Overall, the reviewer considers this a borderline paper with a slight tendency towards recommending acceptance.
Review #3
- Please describe the contribution of the paper
The paper presents a novel collaborative spatial and spectral neural clustering framework called CS2C, aimed at segmenting organelles from volumetric electron microscopy (VEM) images. This approach integrates deep spectral clustering and spatial clustering to enhance the accuracy of organelle segmentation, particularly when dealing with noisy microscopy images. By utilizing features extracted through masked autoencoders, CS2C successfully achieves multi-class segmentation of organelles and outperforms existing state-of-the-art algorithms across multiple benchmark tests.
- 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.
1) The innovative dual-branch clustering design leverages the strengths of both spatial and spectral clustering, enhancing processing outcomes. 2) Experimental results confirm its superior performance in organelle segmentation, particularly its robustness against noise. 3) The provision of open-source code facilitates reproducibility and accessibility for further research.
- 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) The implementation may involve high computational complexity, requiring robust hardware support. 2) The experiments were conducted using specific datasets, which may limit the model’s generalizability to real-world multi-center data. 3) Further exploration is needed to maintain performance across various types of microscopy images.
- 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 submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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
The reviewer is curious about a larger-scale multi-center dataset to validate the proposed 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.
(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 organization and writing of this submission as well as the interesting idea in the manuscript for dual-branch are the major factors.
- Reviewer confidence
Not confident (1)
- [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
Author Feedback
We thank the reviewers for their efforts in reviewing our paper and their constructive comments. @R1, R2, R3: Datasets. We evaluated the proposed CS2C on public BetaSeg with the same dataset settings as SOTA VEM segmentation methods [7, 20]. As pointed out by R2, our approach leverages collaborative neural spatial and spectral clustering for efficient organelle segmentation, which is general and not data-specific. We believe that our method is promising for organelles/species segmentation in a variety of real-world microscopy data, to be explored in future work. @R1: Cluster number. As shown in Table 2, the segmentation performance is insensitive with moderate cluster numbers. In experiments, we set the cluster number to 14 according to the number of organelle types empirically. @R1: Spectral embedding. The spectral embedding via eigendecomposition of graph Laplacian encourages similar nodes to be close in the embedded space, where the nodes of the same structures are aggregated together. Fig. 1 visualizes four sampled spectral bases of VEM images in the “Spectral Embeddings” box. @R1: Comparison of baselines with superpixels. We have compared our approach to SOTA deep clustering methods of FastDGC [5], SDCN [3], and CCGC [21], as well as classical kmeans [13] and spectral clustering [17] on superpixels as shown in Table 1 and Fig. 2. To ensure a fair comparison, we use the same superpixel features and graphs for compared clustering methods [5, 3, 21, 13, 17] as our approach. Our method outperforms baselines with superpixels in reported metrics (Table 1 and Fig. 2). @R1: Code. We have provided an anonymized link to the source code. @R1: Novelty and improvements over the baseline. We propose CS2C, a novel collaborative spatial and spectral deep neural clustering framework, for organelle segmentation. The pillar of our approach is integrating unsupervised deep spectral clustering and spatial clustering, which enhances a harmony of learned cluster assignments under the spatial and spectral superpixel-wise representation. To the best of our knowledge, we are the first to present the spectral-spatial design for unsupervised deep clustering, mitigating high-frequency perturbations in microscopy images and promoting consistent clustering. We have compared with SOTA methods, including DSM [14], DeepCUT [2], FastDGC [5], SDCN [3], CCGC [21], and MAESTER [20], as well as classical kmeans [13] and spectral clustering [17]. Our method consistently outperforms baselines in reported metrics (Table 1 and Fig. 2). For instance, our method achieves IOU gains ranging from 0.071 to 0.282 over compared methods. @R2: Hyperparameters. We empirically set the hyperparameters to 1 to make the neural spectral embedding loss, the clustering loss, and the spatial-spectral consistency regularization loss the same order of magnitude. @R2: Assessing missing object number. We have evaluated the performance using DSC and IoU to measure the consistency of the predicted segmentation with the ground truth. We agree that assessing how many objects were missed, such as granules and mitochondria, would be helpful, to be explored in our future work. @R3: Complexity. In addition to the feature extraction module, we incorporate a lightweight three-layer GCN to approximate spectral bases, a two-layer MLP for clustering, and trainable centroids with a dimensionality of 14x384. The training and testing are on a PC with an NVIDIA RTX 2080Ti GPU only. Note that the GCN-based spectral embedding module reduces the eigen-decomposition complexity from O(n_s^3) to O(n_s(3b^2 + lb+lu)), where b, l, u denote the bandwidth of the graph matrix, the feature channel number, and the spectral bases number. In the inference, the feature extraction, spatial branch clustering, and spectral branch clustering of a 1082x545 image take 2.714 seconds, 0.005 seconds, and 0.002 seconds, respectively.
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.
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