List of Papers Browse by Subject Areas Author List
Abstract
Multi-modal MRI is essential for liver tumor delineation but is often incomplete owing to protocol heterogeneity and structured group-wise modality missingness. We address this setting under a clinically grounded acquisition model, where T2-weighted imaging (T2WI) serves as a permanent anatomical anchor, while dynamic contrast-enhanced phases (pre-contrast C-pre, arterial C+A, venous C+V, and delayed C+Delay), diffusion-weighted imaging (DWI), and Dixon sequences (in-phase InPhase and out-of-phase OutPhase) are available as atomic modality groups. To achieve robust segmentation across all valid permutations, we propose KiD-Seg (Kinetic-Disentangled Contrastive Learning), a unified framework with three key components: (1) Asymmetric Anchor-Guided Contrastive Learning, which aligns auxiliary modalities to the high-fidelity T2WI embedding to reduce latent degradation from distortion-prone inputs; (2) Differential Kinetic Disentanglement, which encodes phase-to-baseline hemodynamic differences to enhance tumor–background separability; and (3) Hierarchical Atomic-Group Fusion with completeness-aware self-distillation, which enforces semantic consistency across sparse and complete modality sets. On the 498-case LLD-MMRI dataset (8 modalities/case), KiD-Seg achieves the best overall median DSC (0.781) and the best full-modality median DSC (0.802), demonstrating robust segmentation performance under both incomplete and complete multi-modal inputs. The source code is publicly available at https://github.com/AQ-MedAI/KiD_Seg.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2415_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/AQ-MedAI/KiD_Seg
Link to the Dataset(s)
N/A
BibTex
@InProceedings{KhoHee_KiDSeg_MICCAI2026,
author = { Khor, Hee Guan AND Zhang, Dong AND Guo, Siyan AND Chu, Tianhao AND Wang, Junyi AND Bai, Xiaoyu AND Shi, Yinghong AND Lu, Le},
title = { { KiD-Seg: Kinetic-Disentangled Contrastive Learning with Anchor Guidance for Incomplete Multi-Modal Liver Tumor Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16884},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The authors introduce KiD-Seg, a novel framework for liver tumor segmentation under group-wise incomplete multi-modal MRI. KiD-Seg uses T2WI as a fixed anatomical anchor and organizes other modalities into structured groups. It combines asymmetric contrastive learning to align modalities, kinetic disentanglement to capture contrast dynamics, and hierarchical fusion to handle missing data consistently. Experiments on the LLD-MMRI dataset show robustness to missing modalities while maintaining competitive performance in all settings.
- 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 addresses a realistic and important challenge (segmentation under group-wise missing modalities in multi-modal MRI), well aligned with clinical acquisition constraints.
- The combination of asymmetric contrastive learning (AGCL), kinetic disentanglement (DKD), and hierarchical group-wise fusion (HGF) is original and thoughtfully designed to tackle complementary aspects of the problem.
- The method explicitly models realistic missingness patterns
- Compared to state-of-the-art, KiD-Seg achieves the best overall median DSC and performs consistently well across across a wide range of modality combinations, including full multi-modal inputs.
- 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 clarity on the missing modality training scheme
- Unclear balancing of the different loss terms
- Insufficient details on dataset and related 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 submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- 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?
/
- 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
The proposed segmentation method is of very high interest for the medical community. Please find below some minor comments which could be taken into account to further improve the paper.
1- The use of phase-wise difference maps in the kinetic disentanglement module assumes accurate spatial alignment between the dynamic contrast-enhanced phases (e. g. , C-pre, arterial, venous, delayed). However, the manuscript does not explicitly state prior to Sect. 3.1 whether these images are registered. Since even minor misalignments could significantly impact voxel-wise subtraction and the resulting kinetic representations, this point should be clarified earlier, in Sect. 2.
2- The rationale for combining both segmentation branches from the base stream and the fused segmentation branch could be better clarified. In particular, it is not entirely clear why the model does not rely solely on the fused branch with HGF and DKD cues, which appears to be the main prediction pathway. This design choice seems not fully justified.
3- It is somewhat difficult to fully understand how the model behaves when modality groups are missing. In particular, the interaction between the student pass (with sampled missing groups) and the parallel teacher pass (with a more complete modality set) is not entirely clear from the current description. A schematic illustration of this process, especially showing how different modality combinations are handled and how knowledge is transferred between teacher and student, would greatly improve clarity.
4- The final training objective combines multiple branch-specific losses and regularization terms, but it is unclear whether any weighting is applied to balance their respective contributions. Given the diversity of objectives (segmentation, reconstruction, contrastive losses, and self-distillation), the absence of explicit weighting coefficients raises questions about their relative influence during training.
5- While the LLD-MMRI dataset includes up to 8 modalities and voxel-wise annotations for 8 liver tumor classes, it is unclear whether the annotations used in this work are the ones defined with respect to T2WI. In addition, the nature of the tumor classes should be clarified: do they correspond to primary liver tumors (e. g. , HCC), secondary lesions (e. g. , metastases), or a mixture of both?
- 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?
- Clinically relevant problem
- Well-motivated framework design
- Innovative methodological contributions
- Robustness to missing modalities
- Strong overall performance
- 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.
- Combining asymmetric contrastive learning (AGCL) with kinetic disentanglement (DKD) and hierarchical group-wise fusion (HGF) in the scope of liver MR scan analysis remains relevant and innovative.
- The explanation provided in the rebuttal regarding the projection of voxel-wise kinetic features into a normalized latent space is now clearer.
- Author should really mention the different labels: hepatocellular carcinoma, intra-hepatic cholangiocarcinoma, metastases, cysts, hemangioma…
- Compared to state-of-the-art, KiD-Seg achieves the best overall median DSC and performs consistently well across a wide range of modality combinations.
Review #2
- Please describe the contribution of the paper
Proposed a framework KiD-Seg for liver lesion segmentation. To solve the following problems: 1.missing sequences in liver MRI; 2.Dynamic sequences are not treated as related. Their proposed model combined three different components: 1.Asymmetric Anchor-Guided Contrastive Learning; 2.Differential Kinetics Disentanglement; 3.Hierarchical Atomic Group Fusion. Their experiments showed imporvements on metrics by using their methods. 4.Using self-distillation to bridge the gap between incomplete/ complete inference.
- 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 could be applied to general multimodality (not only limited to MRI). The idea of using T2WI as a fixed anchor for contrastive learning is new. This is a clinically valid assumption to make. Using contrastive learning for aggregating information for dynamic contrast enhanced is 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.
1.Some of the formulation in the paper are not inmediately clear. For example, in section 2.3, “project voxel-wise kinetic features into a normalized latent space using gk(·)” is not super clear. 2.Ablation study somehow shows (non trivial) decrease in average dice score when HGF is added. 3.The t-SNE is not clearly labeled. I’m not sure but it looks like that with only AGCL shows a better result in seperation.
- 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 provide sufficient information for reproducibility.
- 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
In your main figure, “DWI Diffusion” should just be DWI since it stands for Diffusion Weighted Imaging.
- 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?
This paper uses a lot of off-the-shelf methods and put them into a framework that has been published in last year’s conference (DC-Seg). Despite showing improvements in results, I think this paper should be borderline. I might improve the score if in rebuttle, the author provides 1.a better formulation for the “Kinetic Disentanglement for Dynamic Contrast”; 2.anonymized code if possible.
- 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 several questions I asked in the first review, however, the answers do not fully satisify me but it’s ok. They still do not explain the decrease in average dice score very well (I am aware of the increase in median). Their methdology is still vague despite the effort to explain. They do address my t-SNE questions well. But that being said, I am ok with the paper being either rejected or accepted. I put an accept here because I see the merits of this paper, but that would be a really borderline accept.
Review #3
- Please describe the contribution of the paper
This paper introduces a clinically-based framework for segmenting hepatic tumors using T2WI and auxiliary sequences. The results demonstrate improved performance under data-scarce conditions while preserving robust multimodal outcomes.
- 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 most significant strength of this paper lies in its innovative methodological design, which synergistically integrates multiple modules. This joint formulation enhances the system’s overall coherence and performance. Furthermore, the authors conduct a rigorous comparative evaluation, meticulously examining the model’s behavior both with and without the proposed modules. This detailed analysis provides clear evidence of each component’s contribution and effectiveness, underscoring the robustness of their approach.
- 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 introduction would be strengthened by including specific, quantifiable final results to provide a clearer picture of the paper’s achievements upfront. The manuscript currently lacks a dedicated section or explicit discussion of the study’s limitations and potential avenues for future research. Given the complexity and number of modules presented, a comparative analysis of the model’s runtime and computational resource requirements would be valuable. This would offer a more comprehensive understanding of its practical applicability. While the core ideas are present, the manuscript would benefit from more rigorous editing by a native English speaker proficient in scientific writing. Reliance on AI for language refinement may not fully capture the nuances required for high-impact academic publications.
- 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.
- 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 introduces an interesting methodological design, but it falls slightly below the acceptance threshold due to several key factors: While the paper notes the integration of multiple modules, it does not provide a clear analysis of the computational complexity, including runtime and memory requirements. This omission makes it difficult to fully assess the practical efficiency and scalability of the proposed framework, especially in resource-constrained environments. The manuscript would be significantly strengthened by a dedicated section outlining the study’s limitations and proposing concrete avenues for future research. This would demonstrate a more thorough understanding of the work’s scope and potential for advancement. While the core ideas are present, the overall English quality could be improved. A more rigorous editing process by a native English speaker proficient in scientific writing is recommended to enhance clarity and precision. While the paper presents a novel approach to hepatic tumor segmentation and shows promise, its critical omissions and areas for improvement prevent it from being widely accepted. The identified weaknesses, particularly the lack of computational analysis and refined language, are significant enough to place it marginally below the acceptance bar.
- 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.
Under identical hardware and resolution constraints, the authors demonstrated that KiD-Seg achieves its performance gains with a highly minimal overhead footprint: adding only +0.23M parameters, +0.05G FLOPs, and +9.8 ms of latency. Furthermore, training memory increases by a manageable +0.88GB while inference memory remains entirely unchanged, confirming its immediate readiness for real-world clinical translation. The authors have shown commendable scientific rigor by providing this quantitative complexity analysis, clarifying statistical anomalies, and committing to an open-source code release.
Author Feedback
We thank the reviewers and meta-reviewer for recognizing the clinical relevance of group-wise incomplete liver MRI. We address the major concerns below.
-Novelty w.r.t. DC-Seg (R2/meta-review). DC-Seg is our closest missing-modality baseline, but it mainly treats modalities as exchangeable inputs for generic invariant/specific disentanglement. KiD-Seg is instead derived from liver MRI protocol logic: T2WI serves as a stable anatomical scaffold, DCE phases are interpreted through C-pre-relative enhancement evolution, and DWI/Dixon offer auxiliary protocol-dependent evidence. Thus, (1) AGCL performs asymmetric one-way alignment toward the T2WI anchor with stop-gradient protection; (2) DKD models arterial/venous/delayed enhancement relative to C-pre, capturing tumor kinetics rather than independent phase appearances; and (3) HGF performs atomic group-wise fusion/dropout to match clinically plausible missing patterns. KiD-Seg is a protocol-aware, anchor-guided, kinetic-aware formulation rather than a direct extension of DC-Seg.
-DKD, g_k, and branch rationale (R1/R2/meta-review). “Projecting voxel-wise kinetic features using g_k” means that after phase-difference encoding and cross-phase pooling, z_k(x_i) denotes the kinetic feature vector at voxel x_i; g_k maps it to the contrastive embedding dimension, followed by L2 normalization. The supervised contrastive loss is computed over sampled voxels, using voxel labels to define positive/negative sets. The base segmentation stream is not an alternative inference path; it provides anatomy-guided supervision under incomplete inputs. Since HGF/DKD cues are injected into the fused branch as gated residual evidence and depend on available groups, supervising only the fused branch may under-constrain the base decoder when auxiliary groups are absent. The base stream preserves a discriminative anatomical representation, while the fused branch integrates multi-scale HGF/DKD cues and remains the main prediction path.
-HGF ablation and t-SNE (R2/meta-review). Table 1 reports mean±std with median in parentheses and selects best results by median. Adding HGF improves median DSC from 0.773 to 0.781 and reduces mean/median HD95 from 15.200/5.099 mm to 8.737/4.123 mm. Although mean DSC changes from 0.681 to 0.670, HGF improves typical-case performance and boundary robustness rather than every statistic. For t-SNE, the intended message is both tumor/background discrimination and cross-combination alignment. In panel (a), Full KiD-Seg yields a clearer tumor/background boundary than AGCL alone. In panel (b), stronger color-wise separation indicates stronger dependence on available modality patterns, not better modality-invariant tumor features. Full KiD-Seg instead reduces intra-class tumor variance across modality combinations, which is desirable for incomplete-modality segmentation.
-Reproducibility and dataset details (R1/R2). T2WI is a representation anchor, not an assumption that labels are defined in T2WI space; all modalities are brought into a common registered space before training/evaluation. The eight labels are the dataset-provided liver lesion/tumor categories and represent a mixed lesion setting rather than exclusively primary tumors or metastases. No additional scalar loss weights are used beyond class weights in WCE and the stated temperatures; all Eq. (10) terms use unit weights after normalization. We have prepared anonymized code for release upon acceptance.
-Computational cost and limitations (R3/meta-review). Under the same input resolution and hardware, KiD-Seg improves median DSC by +0.017 and reduces median HD95 by 0.877 mm over DC-Seg, with +0.23M parameters, +0.05G FLOPs, +9.8 ms latency, +0.88GB training memory, and unchanged inference memory, indicating a favorable cost-benefit tradeoff. We will include details and discuss limitations/future work, including T2WI dependence, registration sensitivity, and single-cohort validation.
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 paper introduces a framework for liver tumor segmentation under group-wise incomplete multi-modal MRI. The reviewers agree that the problem is highly clinically relevant and commend the innovative integration of Asymmetric Anchor-Guided Contrastive Learning (AGCL), Differential Kinetics Disentanglement (DKD), and Hierarchical Atomic Group Fusion (HGF). Reviewer 1 is highly supportive (Score: 5), noting the strong overall performance. However, Reviewers 2 and 3 lean towards rejection (Scores: 3 and 3) due to a lack of clarity in methodological formulations, missing computational analyses, and presentation issues.
Points to address in the rebuttal
1.R2 raised concerns about the methodological novelty. Specifically, R2 noted that the framework appears to combine several off-the-shelf methods and may be close in spirit to previously published segmentation frameworks (DC-Seg). The rebuttal should explain more clearly how this work positions with the literature.
2.Reviewers requested clearer technical explanations of core modules. R2 found the formulation in Section 2.3 unclear, especially the statement about projecting voxel-wise kinetic features into a normalized latent space using $g_k$. Reviewer #1 also questioned the rationale for combining segmentation outputs from the base stream and the fused branch, asking why the model does not rely solely on the fused pathway. These design choices should be better justified.
3.Some experimental results and ablations need explanation. R2 noted that the ablation study appears to show a non-trivial decrease in average Dice when HGF is added, which seems counterintuitive. R2 also mentioned that the t-SNE visualization is insufficiently labeled and may even suggest better separation with AGCL alone. These observations should be clarified.
3.Reviewers requested stronger reproducibility and dataset details. R1 asked for clarification on whether annotations are defined with respect to T2WI, and on the exact nature of the eight tumor classes (primary tumors, metastases, or mixed lesions). R1 also asked whether explicit weighting coefficients are used to balance the multiple loss terms. R2 additionally requested anonymized code.
4.R3 was overall positive about the integrated design and comparative evaluation, but still identified practical completeness issues that should be addressed. In particular, R3 requested analysis of computational complexity, runtime, and memory requirements given the number of modules in the framework, as well as a clearer discussion of study limitations and future work. R3 also noted that the manuscript would benefit from stronger language polishing to improve clarity.
- 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.
The authors satisfactorily addressed reviewers’ comments in their rebuttal.
Please ensure that the requested changes (computational times, editing, open-source code release) are performed in your camera-ready version.
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.
After considering the rebuttal and updated reviews, I recommend acceptance. All three reviewers now support accepting the paper, although some concerns remain. Overall, the paper presents a relevant and useful contribution, and the rebuttal has addressed the main concerns. The authors should further improve the methodological description, clarify the reported metrics and ablation results, and follow through on the promised code release in the final version.
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.
The concerns regarding the paper have been largely resolved, and all reviewers agree to accept the paper.
