List of Papers Browse by Subject Areas Author List
Abstract
Detecting small lesions in brain Magnetic Resonance Imaging (MRI) is an important task for clinical early diagnosis, yet their subtle size and appearance restrict the feasibility of both manual annotation and automated segmentation. In addition, current reconstruction-based Unsupervised Anomaly Detection (UAD) methods try to model the global normal distribution, while suffering from the submerged local details and domain shifts between training and testing data. In this work, we propose \textit{FluxCut}, a novel two-stage unsupervised framework that shifts from the global reconstruction mechanism to local geometric analysis. In the first stage, our network learns a general representation of small lesions that maximizes the flux measurement, the resulting flux map robustly enhances the small lesion locations. Then, we extract the localized patch features guided by this map and employ a learning-based graph-cut for fine-grained segmentation.This flux-guided segmentation paradigm encourages the network to focus on local discriminative features, preventing the subtle signatures of small lesions from being overwhelmed by background noise during global reconstruction-residual optimization. In our experiments, we compare with current state-of-the-art UAD methods in multiple 3D datasets to demonstrate that our method achieves significant improvement in all datasets.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3230_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{DenZhi_FluxCut_MICCAI2026,
author = { Deng, Zhiwei AND Shi, Yonggang},
title = { { FluxCut: Shifting from Global Reconstruction to Local Geometry for Unsupervised Small Lesion Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16880},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The authors present a method to segment small brain lesions (such as WMH) combining a representation of the small lesions with a graph-cut approach.
- 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 method enhances the detection of small lesions
- the datasets are public and the code is mentioned to be available
- the method seems to
- 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.
- there are no statistical analysis in the results
- in the experiment part, some methods are missing (such as nn-unet)
- the HD95 and F1 score would have been an interesting metric to add
- maybe some references on classical graph-cut methods are missing
- the results seems lower on MSSEG-2 for example than in the official leaderboard
- 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.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- 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/2026/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 — marginally below the acceptance threshold, but would not mind if 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 suffers il the result/validation part from a lack of comparison with some methods such as https://papers.miccai.org/miccai-2024/288-Paper0920.html
- 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 authors answered the questions/remarks the e had. If they make the corresponding modifications il can be a good article.
Review #2
- Please describe the contribution of the paper
The paper describes an unsupervised method for detecting and segmenting MS lesions in MR images. In contrast to previous unsupervised methods, lesion detection is not treated as an outlier detection problem, by generative modelling the normal brain appearance and treating pathology as a deviation from the normal. Instead the authors propose a 2-stage approach similar to R-CNNs consisting of a detection phase and a segmentation phase. In this first stage, a U-Net is trained to detect the boundaries of lesions, from which the center of a lesion can be derived. Then, a ROI is cropped around each lesion center and a trained U-Net is applied to accurate segment the patch.
The main contribution of this work is a framework that enables the training of both networks without the need of ground truth labels. The network that detects lesion boundaries can be trained without annotating lesion boundaries, by optimizing a flux measure instead, which captures important properties of small lesions. By the same token, the segmentation network can be trained without segmentation masks. Instead, features from the backbone network are used to construct an affinity matrix and the segmentation network is trained to minimize the correlation clustering loss.
The proposed method was evaluated on 3 public datasets and compared against state-of-the-art unsupervised segmentation methods, with the results showing that the proposed method can outperform competing unsupervised methods across all datasets and metrics.
- 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 an alternative formulation for unsupervised MS lesion segmentation, by proposing a method that can be applied like a supervised method, but trained without lesion annotations available, thereby allowing the training on large unlabelled datasets.
Instead of requiring laborious work to annotate lesions, the proposed flux loss captures the essence of how lesions are expected to look like namely that the gradients of all boundary voxel should point to the center of the lesion and train the network against this hypothesis. What makes this approach particularily interesting for a wide range of applications and future research, is that this approach can naturally be combined with supervised learning in cases where labels are available to further improve accuracy. Although not explored in this paper, the flex estimation head essentially predicts the distance to lesion boundaries, which can be directly derived when lesion annotations are available. This enables the natural combination of a small corpus of labelled images with a large corpus of unlabelled images.
The same holds true for the segmentation network. The clustering loss can be used to train a network that essentially performs a voxel-wise classification of the image patch without ground truth. But the network could also be trained when reference segmentation are available thereby combining the usage of laballed and unlabelled data for training.
I think the general framework has a lot of potential to spark research into other loss functions that can be used to formally describe the properties of target pathologies in order to train “supervised” networks on unlabelled data or perform the joint training on labelled and unlabelled data.
- 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.
Unsupervised approaches are inherently difficult to tune for specific edge cases, because the definition of the target is very generic. While supervised methods can relativly easily be tuned by simply adding more training cases that cover failure cases of the network, tuning an unsupervised method requires tuning of the proposed loss functions, which changes the method entirely. However, since this method can be combined with supervised approaches, this may prove as a way to cover failure cases, while otherwise not requiring large amounts of labelled data to train an initial model.
At the first read, I found the method difficult to understand. The explanation of the FEH left the impression that it is trained to estimate the distance to the lesion boundaries, which is true, but sounds like lesion annotations are required to estimate the target distances to train against. It becomes clearer later when the loss function is defined that the computation of the loss actucally does not require target distances and therefore no lesion annotations.
- 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.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- 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/2026/en/REVIEWER-GUIDELINES.html
I think R in equation 1 should be \mathcal{R}.
- 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 proposed approach is very noval in the context of MS lesion segmentation, yet it has the potential for combination with existing supervised methods to even further improve results. I think the work will inspire and/or motivate further research in the area of dedicated loss functions for unsupervised training and for the combination of supervised with unsupervised methods. I think this is crucial to cover failure cases, which are difficult to tune for using unsupervised methods, and to cover the large variability of lesion pathology, and image appearance, which is difficult to cover by supervised methods due to the intensive labelling effort.
- 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.
I think the authors addressed the review comments well. They clarified how the proposed method differs from previous reconstruction-based UAD methods, how the patch extraction does not imped the methods ability to detect small lesions and why no comparison to supervised methods was made. Although the last one would have been nice just to see the gap between unsupervised and supervised methods.
Review #3
- Please describe the contribution of the paper
This proposes FluxCut, a new two-stage unsupervised anomaly detection (UAD) framework that shifts the paradigm from global intensity reconstruction to local geometric modeling, better preserving subtle small lesion details. It introduces an adaptively weighted flux estimation module to robustly locate small, low-contrast lesions by enhancing boundary gradients while suppressing false positives from unilateral edges. It integrates an unsupervised graph-cut segmentation head using correlation clustering and Chan-Vese regularization to generate fine-grained, accurate lesion masks without any manual annotations. It validates the method on multiple 3D brain MRI datasets (MSLUB, MSSEG-2, BrainMetShare) and shows consistent, significant improvements over state-of-the-art reconstruction-based UAD methods for small lesion segmentation.
- 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.Novel paradigm shift from global reconstruction to local geometric modeling 2.Innovative adaptively weighted flux estimation (AWF) for small lesion localization 3.Effective two-stage “localization-then-segmentation” unsupervised pipeline
- 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.Regarding reconstruction, the backbone network training in the first stage essentially relies on the reconstruction head and the MSE reconstruction loss for feature constraints. The so-called “non-reconstruction paradigm” is completely false and misleading academic expression. Its underlying logic of feature learning has not yet broken away from the framework of the reconstruction paradigm. The so-called “paradigm shift” is completely invalid. 2.The main task of the paper is small tumor segmentation. However, in the abstract, the authors state that they are comparing Unsupervised Anomaly Detection (UAD) methods, rather than some image segmentation methods, which leads to some confusion. 3.The performance of the method is entirely dependent on the positioning accuracy of the first-stage flux graph, but no remedial mechanisms have been designed for positioning failure: The local segmentation in the second stage is entirely based on the lesion patches extracted from the flux graph. Once the flux graph fails to locate low-contrast microlesions or misses them, the subsequent segmentation cannot be carried out at all. However, the paper only proves the necessity of the flux head through ablation experiments, without analyzing the proportion of cases where the flux graph fails to locate or misses, nor verifying the method’s detection ability for <3mm clinical core microlesions. 4.The core design is fundamentally contradictory to the goal of small lesion segmentation: The extracted lesion patches are pooled to a fixed resolution of 20×20×20.For microlesions smaller than this resolution, the pooling operation will directly dilute and erase the weak features of the lesion, instead exacerbating the loss of information of small lesions, which is completely contrary to the claim of “preserving the details of small lesions” made by the paper. 5.The core module of the method has no breakthrough innovation: Adaptive Weighted Flux (AWF) only makes simple directional weighted improvements to the existing flux calculation, without any theoretical breakthrough; No specific theoretical innovation has been proposed for small lesion segmentation. 6.In Section 2.1 of the paper, it is stated that the input is 3D images. However, in Figure 1, it is a 2D image, and the author’s display of the segmentation results is also a 2D image. 7.There is a lack of corresponding labels in Figure 1, such as the image input being I. 8.References 14 and 15 are the same. Please check the citation process.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- 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/2026/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.
(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?
The thesis fails to effectively highlight its originality and the structure is rather chaotic. The backbone network training in the first stage of the method essentially relies on the reconstruction head and the MSE reconstruction loss to impose feature constraints. The so-called “non-reconstruction paradigm” is completely false and misleading academic language.
- 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 failed to effectively address the critical concerns regarding the risk of flux map localization failure and core methodological novelty. Most of the key issues were either avoided or not resolved in a substantive manner.
Author Feedback
We thank the AC and reviewers for the comments. We address the concerns below.
1) Reconstruction/paradigm shift. We agree that “non-reconstruction paradigm” was imprecise, since stage-1 training uses a reconstruction head and MSE loss. We will remove this wording. However, using reconstruction as an auxiliary feature constraint differs from reconstruction-residual-driven UAD. FluxCut does not use |I−Î| as final anomaly evidence; reconstruction regularizes the backbone, while AWF drives lesion localization and graph-based local segmentation produces the final mask. Thus, the intended shift is from global reconstruction-residual anomaly mapping to reconstruction-regularized, flux-guided local segmentation, not from reconstruction to reconstruction-free learning.
2) Method clarity. We agree that the original organization made the flux map difficult to follow. We will move the flux-map definition and flux evaluation before the final segmentation framework.
3) Fixed 20^3 resolution. The concern assumes downsampling/average pooling, which is not our implementation. Lesion-centered candidates are resampled to a canonical 20^3 grid; for microlesion patches smaller than this size, this is upsampling, so lesion signals are not averaged into the background. Instead, resampling increases their effective spatial support for the second-stage network. We will replace “pooling” with “fixed-resolution resampling” and revise “preserving details” to “enhancing effective spatial representation”。 4) Methodological novelty. We clarify that FluxCut’s novelty should not be judged only by whether AWF constitutes a new general theory of flux. The main contribution is the flux-guided, graph-based local segmentation framework for small lesions. FluxCut changes anomaly modeling from global reconstruction-residual thresholding to local geometric localization and segmentation: AWF generates lesion-centered candidates, and graph-based local segmentation produces fine-grained masks within these candidates without voxel-level supervision. To our knowledge, this combination is absent from prior reconstruction-based UAD methods and is specifically designed for low-contrast small lesions.
5) Statistical analysis. We adopted Wilcoxon signed-rank tests on per-case results between FluxCut and the strongest baseline in each dataset. FluxCut shows statistically significant improvements in DSC (p<0.003, *), Precision (p<0.009, **), and AP/AUPRC (p<0.0001, **), indicating that the gains are consistent at case level rather than driven by random variation. The significant Precision and AP improvements further suggest that FluxCut improves segmentation quality by reducing false positives rather than simply over-segmenting.
6) Leaderboard and supervised comparisons. The main comparison uses UAD baselines because FluxCut does not use voxel-level lesion annotations for supervised training. Fully supervised methods such as nnU-Net and the recent MSSEG-2 method suggested by R1 are not directly comparable under the same supervision setting, as they use supervised downstream training with real/synthetic lesion labels. We acknowledge MSSEG-2 leaderboard results can be higher, but they mainly reflect fully supervised settings. We will cite and discuss these methods as supervised references/upper bounds, separated from UAD baselines.
7) Flux failure and multiple patches. The “no patch for stage 2” concern does not reflect our implementation. We use a conservative high-sensitivity flux threshold, connected-component analysis, and margin-expanded boxes, so each volume yields candidate patches. For multiple patches, overlapping candidates/predictions are merged by IoU-based grouping after mapping back to the original 3D space. This post-processing will be clarified in the main text.
8) Minor issues. We will correct duplicated references, add missing labels in Fig.1, and add classical graph-cut reference
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
This is an interesting paper that proposes a paradigm shift. Having said that, I want to highlight that this premise (or a hypothesis) is not entirely true as raised by reviewer #3.This proposal is still benefitting from a reconstruction loss and should be considered as a reconstruction method.
Methodological aspects should also be clarified as the writing gets rather confusing and there are details that are not well or fully explained [Reviewer #2]. As a personal comment, I also agree that the huge emphasis on the flux maps without explaining how they are evaluated until much later leads to confusion. Furthermore, I agree with reviewer #3 that the fixed resolution sounds counterintuitive for small lesions and that there is a rather small methodological novelty for a methodological contribution.
Apart from that there are other concerns about evaluation that I think should be addressed:
- Lack of statistical tests (with such low numbers it is hard to judge how meaningful the improvement is) and lack of other metrics [Reviewer #1].
- Lower results than the official leaderboards [Reviewer #2]
- What happens when the flux does not detect the lesion and there are no patches for the second stage for that lesion? Similarly, what happens when multiple patches are selected for a lesion? [Reviewer #3]
I have only highlighted what I consider major flaws, but I would also like to point out that there are other minor concerns by the reviewers that should also be addressed (e.g. duplicated references).
Even though I have highlighted major flaws, I want to also underscore that the authors received a high score and that is the main reason for my decision. Having said that, I think it is important to address all these major concerns carefully.
- 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.
I would like to start by pointing out that there seems to have been a huge misunderstanding between what reviewer #3 was asking in regards to the “failure modes” and what the authors actually replied in the rebuttal. The issue is not the lack of “any box”, but the possibility of missing some lesions in this first stage. In that sense, the authors have not completely addressed that major concern and their answer could even be seen as dismissive (although I see their point!).
Having said that, I think the authors did a good job of addressing all the other major issues as highlighted by reviewers #1 and #2.I also would like to thank them for clarifying confusing parts and being open to downplay some of the claims or at least rephrase problematic sentences. As long as the authors include all the information provided in the rebuttal on the camera ready version of the paper, I have no major concerns. I am also looking forward to reading the newer version where the flux map explanation has been polished.
Meta-review #2
- 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.
The two-stage flux-guided framework is conceptually interesting and the potential to inspire future work on geometric priors for lesion detection is genuine (as noted by R2). However the UAD framing is misleading and was not resolved in the rebuttal (as flagged by R3). By encoding a geometric prior about lesion appearance into the flux loss, the method narrows its scope to detecting anomalies that conform to a specific geometric model of small lesions, which is fundamentally different from general UAD methods that can in principle detect any deviation from a healthy distribution and be used for screening across diverse pathology types. The current comparison against reconstruction-based UAD baselines does not reflect this scope difference. The contribution should be evaluated against (stronger and more recent – many diffusion-based methods were not included) UAD methods with this distinction clearly stated, and against supervised methods as upper bounds, allowing readers to understand what the method can and cannot detect and where it sits relative to the broader field.
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.
This paper proposes an interesting solution to improve small lesion segmentation. All concerns have been adequately addressed in the rebuttal and notably around the problem of resolution, comparison to leaderboards and statistical testing. I would like to see that paper presented at the conference.