Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Segmentations of hepatic vessels, Couinaud segments, and tumors are crucial for precision liver surgery. Existing methods mostly handle them independently, without exploiting their correlation in the segmentations. We propose UniLiver, a unified model for multi-target hepatic segmentation. Specifically, we introduce Multi-Target Gradient-Conditioned Adapter (MTGCA) to resolve the optimization conflicts. By this new means, we address the morphological variations in different hepatic targets and allow the shared encoder to serve each target with minimized mutual degradation. In addition, to handle vast scale and shape disparities among liver anatomical data, we design Target-Aware Spectral-Spatial Encoding (TA-SSE) to modulate spatial positional cues as per target geometric demands. Further, we present VCT Sequential Denoising (VCT-SD) to progressively refine predictions following the anatomical hierarchy from vessels to segments and tumors. Evaluated on the composite LiTS and MSD8 datasets, UniLiver achieves consistent improvements over state-of-the-art methods for specific hepatic targets, confirming its compelling performance of jointly modeling the anatomical dependencies for precision liver surgery. Code is available at https://github.com/YueQiu0911/UniLiver.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1416_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{QiuYue_UniLiver_MICCAI2026,
        author = { Qiu, Yue AND Qiu, Shi AND Heng, Pheng-Ann AND Fu, Chi-Wing},
        title = { { UniLiver: Gradient-Conditioned Unified Model for Multi-target Hepatic 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

    This paper proposes UniLiver, a comprehensive unified model for multi-target hepatic segmentation, including hepatic vessels, Couinaud segments, and tumors. The core motivation lies in exploiting the inherent anatomical correlations between hepatic structures—where vascular topology physically defines the functional liver segments (Couinaud segments), which in turn serve as spatial containers for potential lesions—a synergy often neglected by existing independent models. The model introduces three key components: Multi-Target Gradient-Conditioned Adapter to resolve multi-task optimization conflicts via gradient-conditioned feature adaptation; Target-Aware Spectral-Spatial Encoding to handle morphological disparities using target-aware spectral-spatial encoding; and VCT Sequential Denoising to progressively refine predictions by following the anatomical hierarchy through iterative cross-attention. Experimental results shows that UniLiver achieves consistent improvements over state-of-the-art methods for specific hepatic targets.

  • 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 design of the VCT-SD module effectively translates directed anatomical dependencies into a technical workflow, which is clinically more rational than treating these targets as independent tasks. 2.The MTGCA module addresses a critical pain point in medical multi-task learning—task interference. By applying orthogonal decomposition to modulate features, it allows the shared encoder to optimize both high-frequency and region-level features without mutual degradation. 3.Experimental results demonstrate significant improvements, notably a 4.53% Dice increase in vessel segmentation on the LiTS dataset. Furthermore, ablation studies for the VCT-SD module show that tumor Per-Case Dice (C-Dice) improved by up to 15.56%.

  • 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 framework involves high-dimensional 3D FFT (TA-SSE), iterative denoising steps (VCT-SD), and per-step gradient projections (MTGCA). The paper lacks a detailed analysis of inference latency and GPU memory overhead, which are critical for clinical applications. 2.The sequential nature of VCT-SD implies that inaccuracies in primary vessel segmentation may propagate downstream, degrading the performance of segment and tumor tasks. The robustness of this dependency chain under severe pathological conditions (e. g. , vessels compressed or invaded by large tumors) has not been sufficiently discussed. 3.Despite excellent quantitative results, the qualitative analysis or visualization regarding how the model utilizes the anatomical hierarchy could be deeper to enhance the interpretability and persuasiveness of the findings.

  • 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?

    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.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    This paper proposes UniLiver, a comprehensive unified model for multi-target hepatic segmentation. And the experimental results shows that UniLiver achieves consistent improvements over state-of-the-art methods for specific hepatic targets.It is usful for the cilinical.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Review #2

  • Please describe the contribution of the paper

    The main contribution is UniLiver, a framework/model for multi-target hepatic segmentation. The hepatic structures of interest are: blood vessels, Couinaud segments and tumors. UniLiver is based on DynUNet and introduces three components in the architecture. The authors claim that this framework leads to SOTA segmentation performance for those hepatic targets.

  • 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 joint segmentation of vessels, Couinaud segments and tumors coupled with anatomical hierarchy is clinically interesting.
    • There are 3 separate architectural contributions/modules introduced: multi-target gradient-conditioned adapter (MTGCA), target-aware spectral-spatial encoding (TA-SSE), and sequential cross-attention denoising (VCT-SD). Each one of them adresses a specific sub-problem (gradient conflict, morphological disparity, anatomical dependency). It is a good point rather than just introducing one modification/tweak of the DynUNet network.
    • Results in table 1 show consistent gains for the 3 targets across both test sets, suggesting the potential of the approach. Also single-vs-multi training is done and highlights the value of joint training: it leads to better performance for vessel delineation notably.
    • Ablation studies in tables 2 & 3 are well done, showing both the benefits of different modules combination and the spectral/spatial branches in TA-SSE module.
  • 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.

    Methodology section:

    • The different modules seem to lack details/more depth for technical implementation, none of this can be found in Section 2 or Section 3.1 for the experimental setups.
    • For MTGCA: What is the gradient bank update rule (EMA, last-batch, full-epoch?). What is the nature of the projection matrix P (random Gaussian, orthogonal)? Where exactly the adapter is inserted in the encoder (every stage? bottleneck only?). Furthermore, MTGCA choice is underjustified: even though there is a gap in the literature with few medical tailored approaches (especially for the liver), there is no comparison against the obvious/well established baselines from the multi-task optimization literature (e.g, PCGrad, GradNorm, CAGrad, Nash-MTL, …).
    • For TA-SSE: How the routing weights r computation is determined?
    • For VCT-SD: How the value of S used at training and inference is determined ? How step embeddings are injected?
    • Figure 1 has a typo error: “Confilct Embedding”.

    Experiments/results section:

    • LiTS dataset ships with liver + tumor masks only and MSD8 dataset ships with liver + vessel masks only. Neither provides Couinaud labels. The paper does not state where the Couinaud and vessel labels for LiTS come from. The same observation goes for MSD8 for Couinaud and tumor labels. Number of cases per dataset is also missing/not described.
    • Reference [9] is nnU-Net from Isensee et al. but the paper also says “we adopt a DynUNet backbone [9]”. DynUNet is the MONAI reimplementation of nnU-Net’s topology with some modifications according to this Github issue: https://github.com/Project-MONAI/MONAI/discussions/3149.So citing the nnU-Net paper for the DynUNet backbone is acceptable but slightly imprecise. Also why choosing this backbone and not sticking with the original nnU-Net framework?
    • Loss weighting is unclear: why those specific cutoffs values?
    • DynUNet backbone is initialized with the pre-trained VesselFM weights. If the comparison baselines included in Table 1 were trained from scratch or initialized differently, therefore any gains associated on the vessel task is partly related to the initialization and not the proposed modules. Besides this not very clear, since section 3.1 states the following: “[…] we train and test all comparison methods with the same experimental settings and configurations.” The initialization schemes are the same or different ? Also not using optimized settings/hyperparameters for each baseline could entail underperformance for some of them, a grid search could have been implemented.
    • No statistical testing: all reported numbers in Table 1 are provided without standard deviations or significance tests. UniLiver Single versus Multi comparisons show tight margins for some targets.
    • In Table 1, for Couinaud the most recent comparisons are APVN and Unetr++ released in 2024, but newer Couinaud-specific methods are cited by the authors such as Fan et al. [6] in the introduction and absent in the table. TotalSegmentator is also not included while being a well established off-the-shelf baseline that supports Couinaud segmentation. For tumors, the nnU-Net result (LiTS G-Dice 74.65) is far below the LiTS leaderboard in that benchmark “nnU-Net Revisited: A Call for Rigorous Validation in 3D Medical Image Segmentation”, Isensee et al. 2024.In addition, SOTA baselines such as MedNeXt are missing.
    • A 80/20 random split is used with no cross-validation performed and only 2 datasets used to claim SOTA results.
    • In Table 2, adding MTGCA for the vanilla backbone reduces tumor C-Dice for the MSD8 dataset. The same reduction is observed in vessel HD95 for the LiTS dataset. The analysis does not discuss those points.
  • 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
    • A code release could have been beneficial for reproducibility, the paper introduces 3 different architectural contributions, which is dense in terms of novelty and none of them are clearly described /implemented.
  • 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 problem described in the paper is well motivated and the 3 introduced modules form a coherent flow. The joint multi-target framing brings novelty within the liver segmentation literature and the consistent improvements observed across the three targets and 2 datasets are suggestive. However, the paper is held back by under-specification of the most novel module (MTGCA), absence of the obvious multi-task-optimization baselines, a confusing pretrained initialization that is not isolated in the ablation, an unclear provenance for the LiTS vessel labels along with limited validation (only 2 datasets used). The work appears novel and the empirical results support the claims of the authors on the specific benchmark conducted, but in its present form the paper does not go in depth on what the following are doing: the gradient mechanism, the spectral encoding, the cross-attention denoising, the VesselFM initialization, or the loss weighting. With a thorough revision and a strong rebuttal supplying the missing details and baselines, I could see this becoming an accept.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Review #3

  • Please describe the contribution of the paper

    1) Proposing the Target-Aware Spectral-Spatial Encoding (TA-SSE) module, which learns spectral filtering in conjunction with multi-scale spatial encoding. This module is inherently target-adaptive: a semantic embedding dynamically modulates positional cues, enhancing spatial awareness for anatomically consistent structures such as vessels, while suppressing it for location-invariant targets like tumors. 2) Introducing VCT Sequential Denoising (VCT-SD), which progressively refines predictions in accordance with the anatomical hierarchy. Specifically, VCT-SD performs multiple refinement steps using directed cross-attention to propagate structural information along the anatomical dependency chain—from vessels to segments, and from both to tumors—enabling each segmentation target to benefit from increasingly refined geometric priors derived from its anatomical predecessors. 3) Conducting both qualitative and quantitative experiments on two widely used liver segmentation benchmarks, LiTS and MSD8, and compare our approach against strong baselines as well as state-of-the-art methods.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    The paper precisely identifies the core challenges in multi-target liver segmentation (vessels, Couinaud segments, and tumors): existing methods predominantly handle these structures independently, overlooking the inherent anatomical dependencies among them. This problem holds significant practical relevance in precision liver surgical planning. The methodological innovation is substantial. The MTGCA module introduces a feature modulation mechanism based on inter-target gradient coherence, dynamically resolving gradient conflicts in multi-task optimization by maintaining a gradient bank and applying orthogonal decomposition, enabling a shared encoder to serve multiple anatomical structures without mutual interference. The TA-SSE module innovatively integrates frequency-domain filtering with spatial positional encoding, dynamically adjusting the weight of positional information through target embeddings, thereby achieving adaptive modeling of morphological disparities across different targets. The VCT-SD module constructs a directed acyclic graph following the anatomical hierarchy (vessels → segments → tumors) and progressively propagates structural priors through directed cross-attention, allowing each segmentation target to benefit from its anatomical predecessors. Furthermore, the experimental design is comprehensive, and the visualizations are highly convincing.

  • 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) Insufficient details are provided for certain methodological components: the specific construction of the random projection matrix P in the MTGCA module is not specified (e.g., whether it employs a Johnson–Lindenstrauss transform or a simple random Gaussian matrix); the implementation details of the orthogonal decomposition, such as how conflicting gradient components are isolated, are not provided; the explicit form of the function ψ(F, e_t, γ(p) in the TA-SSE module is not clearly defined; and the specific architecture of the cross-attention block A_{i → j} in the VCT-SD module is not elaborated. 2) A sensitivity analysis of hyperparameters is lacking. The rationale for selecting key hyperparameters, such as the loss weights \lambda_{vessel} = 1.0, \lambda_{tumor} = 3.0, and \lambda_{couinaud} = 1.0, is not discussed. 3) Although the authors claim that the unified model reduces redundant computation, efficiency metrics such as the number of parameters, FLOPs, training time, and inference time are not reported. 4) The conclusion section does not discuss the potential limitations of the proposed method in scenarios involving extremely small lesions, highly adherent structures, or different imaging protocols.

  • 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.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The paper demonstrates strong novelty and holds certain clinical significance. However, the experimental section requires further completeness, and some descriptions in the methodology part need to be more detailed. If the authors can supplement these aspects, I believe the paper could be considered for acceptance.

  • Reviewer confidence

    Somewhat confident (2)

  • [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 meta-reviewer and reviewers for their careful reading and for recognizing the novelty and clinical relevance of UniLiver. We clarify below several points based on the submitted manuscript and experimental setup. For methodology, Section 2 describes the three proposed modules and their roles. Section 2.1 defines MTGCA through a projected gradient bank, target affinity/conflict signal, and feature-wise modulation in Eq. (1)-(2). In implementation, the adapters are placed at each encoder resolution, including the input-block feature and subsequent downsampled encoder features. The gradient signal is maintained online in a compact projected bank and used as an auxiliary conditioning signal for feature modulation. Section 2.2 defines TA-SSE in Eq. (3), where spectral filtering uses input-adaptive routing weights over learnable frequency filters and combines them with target-aware positional encoding. Section 2.3 defines VCT-SD in Eq. (4) as iterative feature refinement over the directed anatomical graph V -> C, C -> T, and V -> T, with timestep embeddings conditioning the refinement process. We also clarify that VCT-SD is not a hard cascade of binary predictions. It performs feature-level cross-attention refinement, so the anatomical hierarchy acts as a soft inductive bias rather than a deterministic post-processing rule. This distinction is relevant to the concern about possible downstream error propagation from vessel segmentation. Regarding dataset construction, the manuscript evaluates on composite LiTS and MSD8 benchmarks for vessel, Couinaud, and tumor targets. We clarify that we did not assume the original LiTS/MSD8 releases natively contain all target annotations. For Couinaud segmentation, the annotations follow the publicly released Couinaud benchmark from Zhang et al., “Robust and smooth Couinaud segmentation via anatomical structure-guided point-voxel network,” which provides Couinaud annotations for public liver CT datasets including LiTS and MSD8-derived cases. Regarding VesselFM initialization, Section 3.1 states that UniLiver uses a DynUNet backbone initialized with VesselFM weights. The VesselFM comparison in Table 1 uses the same pretrained initialization. More importantly, Table 2 provides the controlled ablation: the vanilla multi-task DynUNet baseline and all UniLiver variants share the same backbone and VesselFM initialization, isolating the effects of MTGCA, TA-SSE, and VCT-SD from pretraining. For loss weighting, Section 3.1 reports the target weights used in training. The larger tumor weight is an empirical balancing choice for the stronger class imbalance and lesion-level variability in tumor segmentation. We acknowledge the reviewers’ points on efficiency metrics, statistical testing, broader multi-task optimization baselines, severe-pathology robustness, and deeper hierarchy-focused visualization. We will further revise our camera-ready paper by taking them into consideration.



Meta-Review

Meta-review #1

  • Your recommendation

    Provisional Accept

  • 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.

    The reviewers agreed that the paper is novel and clinically useful, rating two WAs and one WR. I recommend acceptance of the paper, but hope the authors can further refer to reviewers’ comments and clarify missing details in methodology and evaluation. A release of the source code would be beneficial.



back to top