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

Abstract

Spatial transcriptomics (ST) provides unprecedented opportunities to understand tissue structures. Spatial domain identification serves as a core task to partition tissue regions. Current methods achieve initial success on such task, however struggle to identify accurate domain boundaries and rarely attempts supervised paradigm. To address these challenges, we propose a Label-efficient implicit model with Spatially Adaptive Directional total variation (LiSAD). LiSAD adopts implicit neural representations (INRs) with a hypergraph learning framework to transform discrete ST data into continuous representations while modeling high-order topological relationships for spots. We propose a spatially adaptive directional total variation (SADTV) regularization for accurate boundary identification. Furthermore, we propose a label-efficient optimization strategy that leverages minimal expert prior knowledge to learn refined biological representations. Experimental results on multiple datasets show that LiSAD outperforms current methods by 3-7% in adjusted rand index (ARI), validating its effectiveness. We propose a practical label-efficient paradigm for ST analysis. LiSAD achieves a superior trade-off between annotation economy and diagnostic accuracy, enabling accurate domain identification with minimal expert priors. Code is available at https://github.com/BioMedIA-repo/LiSAD.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5026_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/BioMedIA-repo/LiSAD

Link to the Dataset(s)

N/A

BibTex

@InProceedings{XiaYin_LiSAD_MICCAI2026,
        author = { Xia, Yinghao AND Ji, Xiaoxuan AND Liu, Haoyang AND Guo, Fei AND Feng, Xikang AND He, Liang AND Cui, Hui AND Xu, Shuang AND Su, Ran AND Wei, Leyi AND Jin, Qiangguo},
        title = { { LiSAD: A Label-Efficient Implicit Model with Spatially Adaptive Directional Total Variation for Spatial Transcriptomics } },
        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

    This paper addresses the spatial domain identification in spatial transcriptomics (ST) by transforming ST data into a continuous representation using implicit neural representation (INR) and a hypergraph learning framework with a label-efficient (supervised) scheme. It also uses a total variation-based regularizer for boundary identification. The paper identifies 1.ST data being sampled on a discrete spatial grid, 2.non-sharp transitions between domains in a tissue sample, 3.failing to use at least a limited number of pathologist annotated spots as the deficiencies in existing graph-based ST domain segmentations. This paper used INR to transit from discreet ST to continuous, a spatially adaptive directional total variation (SADTV) regularizer to preserve boundaries, and some pathologist labeled spots.

  • 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 paper is well-written, and almost all the aspects a reader anticipates have been addressed.
    2. The ARI and NMI results with the datasets DLPFC, HBC, and STARmap have significantly improved.
    3. The ablation shows the contribution of individual interventions.
    4. Visualizations in Fig. 2 show domain segmentations much closer to the ground truth.
    5. Fig. 3 shows the contribution of SADTV to resolve thin boundaries.
  • 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 three interventions—neural representation of ST data, a directional total variation regularizer, and some pathologist labeled spots—themselves are not novel.
    2. Does Lipschitz continuity result from the INR itself or is it dependent on the underlying spatial transcriptomics dataset? How essential is this for the overall performance? Also, can SADTV be applied on the reconstructed gene expressions?
    3. As I noticed, M = 30 spots are annotated by pathologists. How should this number depend on the number of domains to be segmented? How are the spots selected for annotation? I have noticed Fig. 2b which shows saturated performance beyond M = 5.
    4. None of the four papers against which the results are compared report numerical results. Therefore, I presume that the results in Table 1 are authors’ replications. This hinders conforming that the results are actually superior. A comparison with papers that report ARI, e. g. , Luo et. al. , STINR: Deciphering Spatial Transcriptomics via Implicit Neural Representation, CVPR 2025, is important.
    5. How would the performance be affected if the a. spatially varying, and b. directional nature in total variation loss formulation are removed?
    6. The first addition in Table 2 is pathologist annotated labels (label efficient). I am curious to know the improvements achievable by INR and SADTV without pathologist annotated labels.
  • 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.

  • 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 SADTV is the no. of nearest neighbors equal to 2k or K? Consider using a different symbol for the annotation dataset as D has already been used for no. of gene expressions. Consider citing the papers in the tables too for easy reference. Please give descriptive caption for tables.

  • 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?

    Considering the strong results and the justifications, I selected “Weak Accept.” However, I still have concerns regarding novelty and the independent verification of results. I would be willing to upgrade my rating if these concerns are adequately addressed.

  • Reviewer confidence

    Confident but not absolutely certain (3)

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

    Accept

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

    The rebuttal answers almost all the questions. Other reviewers too seem to perceive this work as a valuable contribution.



Review #2

  • Please describe the contribution of the paper

    The authors introduce a label-efficient implicit model for accurate boundary identification in spatial transcriptomics. The method integrates hypergraph learning and a variation regularization strategy to model topological relationships, while the overall framework is trained under a supervised paradigm.

  • 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 paper proposes a simple yet effective framework with INR-enhanced hypergraph representation learning to capture high-order topological structures within the tissue microenvironment, and a SADTV regularization strategy to ensures the smoothness within the spatial domain. (2) The incorporation of supervised learning enables efficient utilization of limited expert annotations and demonstrates clear performance improvements. (3) The paper is well written and easy to follow, with clear model diagrams and visualizations.

  • 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 paper basically aggregates some commonly used components, such as hypergraph learning and supervised objectives, and therefore exhibits limited methodological novelty. (2) There appears to be an issue in Equation (5): the numerator in the exponent of the denominator should be “pm”. (3) In Equation (6), the meaning of the “i” preceding ϕi,j is unclear and requires clarification. (4) The data split strategy is not clearly described; a detailed explanation of how the training and test sets are splitted is needed to improve reproducibility. (5) The LiSAD (Full) model shows only moderate improvement compared with Label-efficient + INR baseline, and the ablation study is conducted on a single cohort, raising concerns about the effectiveness and generalizability of the SADTV regularization.

  • 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.

  • 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 presents a technically sound framework with clear performance improvements. However, the methodological novelty is limited, and several details (e.g., equations and data split strategy) require clarification. In addition, the experimental validation is somewhat limited, particularly in terms of generalizability. Overall, both the methodological novelty and the experimental validation require further clarification to fully substantiate the effectiveness of the proposed approach.

  • 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.

    Thank you for the authors’ feedback. My major concerns have been addressed, and I am inclined to accept this manuscript.



Review #3

  • Please describe the contribution of the paper

    This article introduces a label-efficient implicit model for spatial domain identification in spatial transcriptomics (ST). It integrates implicit neural representation and hypergraph neural network to convert discrete ST data into continuous spatial fields. Modeling high-order cellular topological relationships overcomes the limitations of discrete sampling and fixed pairwise graphs. It introduces a label efficient supervision framework. Using minimal expert annotations to guide representation learning greatly reduces annotation costs while improving accuracy, making it suitable for clinical pathology workflows.

  • 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.It combines INR for continuous spatial modeling with hypergraphs for capturing higher-order cell communities. This design better captures subtle spatial changes and complex tissue structures. 2.The proposed SADTV enhances the true domain boundaries while suppressing noise, improving the problem of inaccurate boundaries in conventional total variation and clustering methods. 3.This framework uses extremely sparse expert labels that are consistent with the real pathology annotation workflow. 4.This paper is well-organized, with intuitive illustrations, clear mathematical formulas, and thorough explanations of motivation, methods, and results.

  • 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.SADTV relies on pre-defined K-neighbors, which may still fail in highly irregular or low-quality tissue slices, similar to the fixed neighbor GNN model. 2.The core innovation of this article is not well expressed. On the contrary, they seem to be a loose combination of incremental technology modules, including implicit neural representation (INR), hypergraph modeling, total variation (TV) regularization, and label efficient learning. Further refinement is needed to express innovative points. 3.Some mathematical symbols and expressions lack clear definitions when they first appear, which reduces readability.

  • 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.

  • 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 overall evaluation of this article is based on technological novelty, clinical relevance, contribution novelty, limitations, and writing quality.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [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 author’s explanation and description of the innovation are insufficiently in-depth and fundamental, and fail to reveal the internal mechanisms among the combined modules from fundamental principles. Meanwhile, the illustration of the mechanism of SADTV, one of the core innovations, is rather weak, with limited analysis of its advantages and essential principles. Therefore, based on the current statements in the paper and explanations provided in the rebuttal, the overall contributions of the proposed method remain unclear.



Author Feedback

We thank the reviewers and address the concerns below. 1.Novelty of LiSAD (R1Q1, R2Q1, R3Q2) Application-wise, LiSAD novelly shifts spatial domain identification from unsupervised algorithms to a clinically interactive paradigm guided by expert annotations, which is more suitable for clinical scenarios. Technically, LiSAD introduces INR and hypergraph learning to address discrete sampling and multicellular modeling, respectively, offering a continuous and higher-order topological representation scheme for ST. To address the challenge of boundary identification, LiSAD further designs SADTV as a regularizer tailored for ST data, providing a new formulation for boundary-sensitive modeling in ST. Thus, we believe LiSAD demonstrates sufficient novelty and is not a simple assembly of existing components. 2.SADTV Effectiveness and Ablation Study (R1Q5, R2Q5) As stated in Sec. 3.3, SADTV shows moderate overall gains because it mainly targets boundary regions, which occupy only a small portion of the tissue section. Fig. 3 shows improved boundary recognition. Moreover, removing the adaptive weighting term and the directional term leads to ARI drops of 0.01 and 0.04 on DLPFC, respectively, validating the proposed design. SADTV also shows consistent effectiveness across datasets: removing it decreases ARI by 0.05 on HBC and 0.03 on STARmap, suggesting its potential generalizability to similar clinical problems with ambiguous tissue boundaries. 3.Strategy for selecting expert-annotated spots (R1Q3) Based on clinical practice, expert labels are selected from high confidence spots inside domains rather than ambiguous boundary regions. To simulate this process, we compute each spot’s KNN neighborhood label purity as q_i = (1/K) * sum_{p_j ∈ N_i} 1(y_j = y_i) for each spot p_i, and randomly select M annotated spots from high purity candidates in each domain. Only these selected labels are used for optimization. 4.Dataset split (R2Q4) As stated in Sec. 2.3, only expert annotated spots are used for optimization, while all remaining labels are used only for whole slice evaluation. This follows the common spatial domain identification protocol, rather than a conventional train/test split. 5.Fairness of baseline reproduction and comparison with STINR (R1Q4) Table 1 baselines are reproduced with official code and recommended settings under a unified environment. Our reproduced results are in the same range as those reported in the STAIG supplement, supporting the reliability of reproduction and the fairness of the comparison. For reference, STINR reports an ARI of 0.63 on DLPFC, while LiSAD achieves 0.76.LiSAD still clearly outperforms STINR, confirming its effectiveness. 6.Target of SADTV and Neighborhood Design (R1Q2, R3Q1) We clarify that SADTV is applied to the reconstructed gene expression, as shown in Eq. 6.For predefined neighborhoods, SADTV uses 2K-nearest neighbors to enlarge the local receptive field and adaptive weights to adjust neighbor contributions, mitigating fixed-neighborhood rigidity. 7.Performance without expert annotations (R1Q6) Without expert priors, the label-free variant HAE+INR+SADTV with embedding clustering achieves an ARI of 0.63, higher than STAIG’s 0.61 in Table 1, indicating that the proposed components still improve representation quality without expert supervision. 8.Source and role of Lipschitz continuity in INR (R1Q2) Lipschitz continuity arises from INR-based modeling, as supported by [10]. It smooths and stabilizes the reconstructed data, providing a noise-robust basis for SADTV. The ablation results about INR in Table 2 verify its effectiveness. 9.Definition of M and total annotation number M×S (R1Q3) There may be a misunderstanding about the definition of M. As stated in Sec. 2.3, S is the number of domains, and M is the number of labeled spots per domain. Thus, the total annotation number is M×S, linear in the number of domains. 10.Writing Issues We will correct all issues in notation.



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.

    The reviewers generally liked the clear writing style, the promising empirical performance, and the potential clinical relevance of exploiting the fact that only a limited set of expert annotations is required to identify the spatial domains in a clinical dataset. The main concerns, though, are (1) limited methodological novelty, as the authors are just combining pieces of existing works and it is not clear what their unified new contribution is; (2) the limited ablation results which should have made SADTV (and thus the paper) more convincing; and the potential generalizability of SADTV to other similar clinical problems; (3) label-selection and data-splitting procedures are not clear to us and should be clarified to better reflect the label-efficient setting in the paper; (4) the comparison results are not reported fairly by the baseline models. For instance, the comparison between our results and the reported ones of baseline methods is actually comparing with our reproduction of those baseline results, instead of directly comparing to the reported ones of the baseline. Rebuttal should discuss these questions and address those issues with sufficient explanations, justification, and evidences.

  • 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 rebuttal, R1 and R2 support acceptance, finding that most concerns were addressed, while R3 remains concerned that the innovation and SADTV mechanism are not sufficiently explained. After careful review, the concerns could not block this paper in my opinion. I vote for acceptance.



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.

    The rebuttal addresses several methodological and protocol concerns, including the annotation selection strategy, the absence of a conventional train/test split, baseline reproduction, and additional SADTV ablations across datasets. Although the method remains an integration of existing components and the novelty is moderate, the clarified label-efficient clinical setting and post-rebuttal reviewer balance support acceptance as a borderline/weak accept.



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 paper presents a clinically relevant label-efficient framework for spatial transcriptomics domain identification, with consistently strong empirical performance and clear gains over both unsupervised and label-efficient baselines. The rebuttal adequately clarified most concerns regarding label selection, dataset splitting, baseline reproduction, and SADTV ablations; although novelty and mechanistic explanation remain somewhat limited, the overall contribution and evidence support acceptance.



back to top