List of Papers Browse by Subject Areas Author List
Abstract
Spatial transcriptomics (ST) measures mRNA expression while preserving spatial organization, but multi-slice analysis faces two coupled difficulties: large non-rigid deformations across slices and inter-slice batch effects when alignment and integration are treated independently. We present INST-Align, an unsupervised pairwise framework that couples a coordinate-based deformation network with a shared Canonical Expression Field, an implicit neural representation mapping spatial coordinates to expression embeddings, for joint alignment and reconstruction. A two-phase training strategy first establishes a stable canonical embedding space and then jointly optimizes deformation and spatial-feature matching, enabling mutually constrained alignment and representation learning. Cross-slice parameter sharing of the canonical field regularizes ambiguous correspondences and absorbs batch variation. Across nine datasets, INST-Align achieves state-of-the-art mean OT Accuracy (0.702), NN Accuracy (0.719), and Chamfer distance, with Chamfer reductions of up to 94.9% on large-deformation sections relative to the strongest baseline. The framework also yields biologically meaningful spatial embeddings and coherent 3D tissue reconstruction. The implementation is available at https://github.com/BonianHan/INST-Align.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2051_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/BonianHan/INST-Align
Link to the Dataset(s)
https://zenodo.org/records/14906156
BibTex
@InProceedings{HanBon_INSTAlign_MICCAI2026,
author = { Han, Bonian AND Qi, Cong AND Musialski, Przemyslaw AND Wei, Zhi},
title = { { INST-Align: Implicit Neural Alignment for Spatial Transcriptomics via Canonical Expression Fields } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16889},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper presents INST-Align, an unsupervised framework for spatial transcriptomics slice alignment and reconstruction. The main contribution lies in coupling a deformation network with a shared canonical expression field, enabling joint optimization of spatial alignment and expression modeling within a continuous representation. Experimental results show competitive alignment performance across multiple datasets and demonstrate that the learned representations can support downstream biological analysis.
- 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.Although INST-Align is optimized for alignment (embedding quality is a byproduct rather than the primary target), it still produces spatial embeddings that are useful for clustering analysis.
2.The Spatial Expression Alignment Loss and Jacobian Regularization are biologically motivated and reasonably designed.
- 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 abstract claims state-of-the-art performance, but the results are not consistently strong across all datasets (e. g. , D1 and D2), where existing methods remain competitive or superior. A more balanced discussion and analysis of these cases would strengthen the paper. 2.It is not clearly specified whether all baselines use the same preprocessing (e. g. , adaptive ICP) and data splits. In addition, key hyperparameters (e. g. , in Eq. (6) and (8)) are fixed without sensitivity analysis, raising concerns about robustness and reproducibility. 3.The ablation does not fully disentangle the contributions of individual components, particularly the loss terms in Phase 2 (e. g. , L_recon vs. L_match). Moreover, it is conducted on a single dataset, which limits its generality. 4.While the method is designed to jointly perform alignment and integration, the results suggest that improvements are uneven. In particular, Phase 1 mainly enhances embedding quality, while its impact on alignment metrics is less evident, weakening the claim of strong mutual benefit. 5.Although the method achieves favorable average performance, improvements are not consistently dominant across datasets or metrics. For example, gains in Chamfer distance are not always matched by comparable improvements in OT or NN accuracy, making the practical advantage less clear.
- 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
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.
(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 paper tackles an important problem in spatial transcriptomics and proposes a technically sound framework with biologically motivated components. It achieves competitive performance and produces useful spatial embeddings for downstream analysis. However, the contribution is limited by insufficient evidence: the claimed state-of-the-art performance is not consistently supported, experimental fairness is unclear, hyperparameter sensitivity is not analyzed, and the ablation study is incomplete. The method presentation could also be clearer. Overall, the work is promising but remains borderline due to limited validation and analysis.
- 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 appreciate the authors’ detailed rebuttal and clarification efforts. The responses sufficiently addressed my concerns regarding the motivation of the paired-to-canonical design, component isolation, computational cost, and reproducibility, and the additional explanations are generally consistent with the empirical results presented in the manuscript. While some limitations remain, particularly regarding biological validation and global multi-slice consistency, I believe the work presents a technically solid and interesting framework for joint spatial alignment and integration in spatial transcriptomics. Therefore, I maintain my accept recommendation.
Review #2
- Please describe the contribution of the paper
This paper proposes INST-Align, an unsupervised pairwise framework for spatial transcriptomics (ST) slice alignment, featuring a 2-stage coupled non-rigid deformation network with a shared Canonical Expression Field implemented as implicit neural representation (INR). Slice-specific expression INRs with a shared decoder are first pretrained to establish a common embedding space; the source INR is then removed and the deformation network are jointly optimized with the canonical decoder. Loss includes a geometry-feature matching loss, reconstruction loss, and Jacobian regularization. The paper reports experiments on nine datasets and claims state-of-the-art mean OT Accuracy, NN Accuracy, and Chamfer distance, and unaffected embedding quality.
- 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 motivated by a real and important problem. Multi-slice spatial transcriptomics alignment remains a challenging. The paper’s motivation for coupling geometric alignment with expression-level information is logical and broadly consistent with recent reviews and benchmarking studies in ST. 2.The proposed combination of a shared canonical field and coordinate deformation is novel and technically reasonable. 3.The model architecture is clearly presented. The two-stage fitting (pretraining and joint training of deformation and canonical consistency) is well established and logically organized. 4.The method is evaluated on diverse spatial transcriptomics datasets, and the paper reports multiple alignment metrics, including OT Accuracy, NN Accuracy, and Chamfer distance. In addition, the authors provide a downstream evaluation of biological embedding quality.
- 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 experimental gains are somewhat uneven, and the biological validity of the learned embeddings remains under-validated. Some performance variation is expected given the noisy and unsupervised nature of spatial transcriptomics alignment. However, the downstream analysis in Table 2 shows that INST-Align does not achieve SOTA embedding quality, which weakens the strength of the claimed “joint alignment and integration” formulation. This cannot be fully dismissed as a mere “byproduct”, since the paper emphasizes biologically meaningful embeddings, and getting a good embedding is the final touch to a scientific analysis. 2.The manuscript does not provide clustering visualizations. Although the paper includes a 3D reconstruction example on DLPFC, the visualization is shown only with ground-truth labels, and DLPFC itself exhibits relatively limited slice-wise deformation, making this example less compelling as evidence for the claimed advantages. 3.The central methodology is not fully analyzed. The manuscript does not provide a direct analysis of how the shared canonical field mitigates batch effects, nor does it isolate the contribution of the proposed Spatial-Expression Alignment Loss in the ablation study. This is important because recent benchmarking work suggests that multi-slice ST performance is highly context dependent and that alignment and integration quality should be dissected carefully. 4.Implementation details and computational cost are insufficiently reported. This is a concern because the full training pipeline requires 700 epochs in total (300 + 400), which is difficult to judge the practical efficiency of the proposed method relative to competing alignment approaches.
- 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
1.In Phase 1, two separate ExprINRs are trained for the reference and target slices, respectively, but the target-slice INR is discarded in Phase 2.If the target-slice encoder is no longer used, how does the final “shared” canonical expression field retain target-specific biological information in a principled way? More specifically, what is the conceptual advantage of the proposed design over using only the reference slice to define the ExprINR and then learning deformation from the target slice into that space? 2.In Fig. 2(a), the two slices appear to have noticeably inconsistent orientation after rigid ICP pre-alignment. Could the authors clarify the precise role and necessity of the initial rigid ICP step? 3.Could the authors provide implementation details such as training device, wall-clock time per subject, number of optimization steps, and memory usage? Given that the method requires a 700-epoch two-stage optimization schedule, the computational cost may be substantially higher than that of many competing alignment methods, some of which run in seconds to minutes. I would be open to raising my score to 4 or 5 if the authors can convincingly address the concerns above.
- 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 has a good idea, addresses an important problem, and reports promising results, especially in difficult large-deformation cases. The methodological intuition is sound and the technical design is coherent. However, the paper does not yet validate its central assumption with the level of experimental rigor. In particular, it does not convincingly isolate the benefit of the shared canonical field as the main causal factor, and its biological validation remains limited relative to the strength of its claims.
- 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 addressed most of my concerns I will raise the score to 4. Details; The authors gave a careful and robust rebuttal in response to the question I raised. Most of the confusions have been settled, including how the pre-ICP alignment works, how the canonical space is built, and several issues regarding to the experiment results. These clarifications successfully differentiate the author’s innovations from incremental developments from previous methods or mere overclaiming, thus I would like to accept the paper.
Review #3
- Please describe the contribution of the paper
This paper presents INST-Align, a novel unsupervised pairwise framework for spatial transcriptomics that couples a coordinate-based deformation network with a shared Canonical Expression Field. The core contribution is the joint optimization of spatial alignment and expression integration, utilizing an implicit neural representation to map spatial coordinates to expression embeddings. The authors emphasize that cross-slice parameter sharing of this canonical field regularizes ambiguous correspondences and absorbs batch variation, achieving state-of-the-art performance across multiple datasets.
- 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.
Accurate reconstruction of coherent 3D tissue volumes from serial sections is a critical challenge, as slices undergo complex non-rigid distortions such as tearing and folding during sample preparation. However, existing alignment methods often treat alignment and integration as independent steps. Relying on integration-based methods risks over-correcting and diminishing genuine biological diversity, while geometric methods lack expression-level feedback. Furthermore, biological complexity makes ground-truth correspondences largely unattainable, raising the need for a robust unsupervised strategy. To overcome these limitations, the authors design INST-Align, which employs a two-phase training strategy to establish a stable canonical embedding space before jointly optimizing deformation and spatial-feature matching. Furthermore, the proposed spatial-expression alignment loss combines geometric proximity and canonical-field feature similarity to drive unsupervised non-rigid registration. This design aims to turn alignment into a mutually constrained reconstruction task.
- 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 integration of a deformation network with a shared canonical field represents a highly promising and valuable approach for ST multi-slice alignment. To further elevate the clarity and impact of this strong manuscript, the authors might consider clarifying a few methodological details: 1.The use of an asymmetric weighting scheme in the Jacobian regularization to penalize tissue compression more heavily than expansion is an intuitive and physically plausible design choice. However, since the specific weights chosen directly impact the smoothness of the displacement field, their specific values could benefit from a bit more validation. The paper would be further strengthened if the authors could provide a brief ablation study or visualize the resulting vector fields under different weight settings. This would beautifully illustrate that they indeed preserve biologically plausible cell densities without introducing artifacts. 2.The claim that the spatial-expression alignment loss utilizes an adaptive temperature and EMA scale factors is compelling. While the current two-phase training is a solid baseline, the interpretability of this dynamic matching could be made even more convincing. The authors might consider visualizing the convergence curves of these dynamic scaling factors during the joint optimization phase. Showing that these parameters stabilize naturally over the 400 epochs would wonderfully solidify the paper’s core contribution regarding optimization stability. 3.The model achieves impressive state-of-the-art results across most datasets, particularly in Chamfer distance reduction on large-deformation sections. However, to provide a more comprehensive evaluation, it would be valuable to briefly discuss the potential for error accumulation when performing 3D tissue reconstructions from consecutively aligned slices. Including a short discussion in the conclusion regarding how the current pairwise framework might be extended to include a global multi-slice consistency objective would provide readers with a more objective understanding of the model’s limitations and future scalability. 4.The biologically meaningful embeddings generated by INST-Align are one of its strongest aspects. There is, however, a minor area for expansion regarding the biological validation, which is currently limited to clustering-based metrics like ARI and NMI. Could the authors kindly clarify or evaluate how these embeddings perform on specific spatial trajectory inference tasks or continuous phenotype mappings? Detailing this downstream utility would greatly help readers fully understand the mechanics and ensure the biological relevance of the canonical field.
- 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?
Novelty and Significance (High) Methodological Rigor (Medium) Experimental Validation (Medium) Reproducibility and Transparency (Low)
- 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
Author Feedback
We thank the reviewers and MR. All numbers below are from the submitted manuscript or characterizations of the submitted runs.
1.SOTA claim [R1.1, R1.5, R2.1, MR-1]. The claim is best mean across nine datasets. On D1/D2, top methods differ by at most 0.037 OT and 0.031 NN; on D3 we lead OT (.839) and NN (.849). The advantage is decisive under large deformations: Chamfer 12.6/17.1/11.9 on MB2/MB11/ME vs. 166.2/103.3/12.0 (Spateo) and 173.1/51.7/235.5 (PASTE). Mean OT .702, NN .719, Chamfer 22.6 are best across all baselines. The abstract phrase “the strongest baseline” will be replaced with “PASTE”, consistent with §3.2 and Conclusion.
2.Why paired-then-canonical [R2.Q1, MR-2]. A reference-only ExprINR would bake the reference’s batch signature into the canonical signal; deformation would then drag the source onto reference-specific artifacts. Phase 1’s two ExprINRs share one decoder (§2.2): the decoder must reconstruct both slices from coordinates alone, forcing slice-specific noise into the reconstruction residual and out of the canonical embedding. This is the integration step. The source INR is discardable in Phase 2 because the deformation network maps source coordinates into the canonical frame, where one shared field suffices. The design is the explicit mechanism behind “joint alignment+integration”, not redundant scaffolding.
3.Component isolation [R1.3, R2.3, R3.1, MR-2]. Table 3 dissociates three orthogonal axes: removing Phase 1 collapses embedding (ARI .334->.045) with alignment intact (OT .810->.813); removing Phase 2 degrades alignment (Chamfer .36->.56); removing Jacobian degrades both (Chamfer .49, ARI .287). The Phase-1 ablation pattern reproduces on MouseEmbryo (OT .75->.69, ARI .22->.13), confirming the trend is not dataset-specific.
4.Cost and reproducibility [R1.2, R2.Q3, MR-4]. On MouseEmbryo (~37k cells, single RTX 5080 16GB) INST-Align trains in 95 s – faster than PASTE (232 s) and STalign (435 s), competitive with Spateo. The 300+400-epoch schedule is dataset-agnostic and runs on consumer hardware. All datasets use the identical Dong et al. [11] pipeline; baselines run released code under default init. Core design is fixed across all nine datasets – two-phase training, shared canonical field, spatial-expression matcher, 5:1 Jacobian asymmetry. Only scale/geometry settings vary, and were fixed before final evaluation: robustness in the architecture, not missing sensitivity.
5.Specific items. R2.Q2 (ICP). Rigid ICP removes only translation/rotation; the residual in Fig. 2(a) is exactly the tearing/folding the non-rigid step corrects – expected, not failure. R3.1 (Jacobian). The 5x compression penalty follows the local-invertibility argument of [29]: near-singular compression destroys the diffeomorphism, expansion does not. Table 3’s “w/o Jacobian” row quantifies the effect. R3.2 (tau / scales). In the submitted runs, tau, s_sp, s_ft stabilize within 0-25, 0-10, 12-65 epochs across all datasets – natural early self-calibration. R3.3 (drift). Acknowledged as a current limitation in the Conclusion. R2 / R3.4 (biology). The canonical field is coordinate-conditioned and differentiable; trajectory inference and continuous phenotype mapping are downstream uses of the same trained model, not separate methodology – consistent with Table 2, where our embeddings beat PCA and Seurat despite alignment being primary.
Closing. Each MR ask – inconsistent gains, isolation of contributions, utility of embeddings, reproducibility – is answered from the submitted Tables 1-3 plus characterizations of the submitted runs, in compliance with rebuttal policy. R1 already recommended Accept (5, very confident). R2’s Q1/Q2/Q3 are addressed in §2/§5/§4 and R3’s asks in §3/§5; on R2’s own terms (“open to raising my score to 4 or 5”), we ask R2 and R3 to update their scores.
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.
It is an interesting paper that presents an important problem and proposes a coherent method with a novel coupling between non-rigid deformation and a shared canonical expression field. However, the results do not quite convincingly support the claims. The performance gains are inconsistent across datasets and metrics. It is unclear how much of the gain results from joint alignment and integration, especially from using a shared canonical field and spatial-expression alignment loss. We believe that more experiments are needed to better demonstrate the utility of the learned embeddings. Reproducibility needs more information such as fairness of preprocessing, the sensitivity to hyper-parameters, the run time, and any other relevant information for replicating the experiments. Please carefully address these points with clear evidence and additional ablations/analyses, and make the manuscript more balanced.
- 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 maintains Accept after the clarifications on design, ablations, cost, and reproducibility, while R2 raises the score to Accept after the authors addressed concerns about pre-ICP, canonical space, and experimental interpretation. R3 originally rates weak reject but did not submit final rating. After review, I think the concerns would not block this paper. 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.
This work proposes a registration approach for spatial transcriptomics from different sections. The developed approach makes use of implicit neural representations to model the spatial deformation as well as mRNA expression.
While there was mixed review enthusiasm before the rebuttal, reviewer sentiment was largely positive after the rebuttal. Reviewers appreciated the tackled problem and the average gains obtained by the proposed method compared to competing approaches. There are some remaining concerns regarding the quality of the obtained joint embeddings (as performance is lower than for specialized methods) and the ability of the approach to extend to multi slice alignment (the dreaded banana effect comes to mind for 3D reconstructions from pairwise 2D registration) but overall this paper provides an interesting approach for a challenging task. Hence, in my opinion, it is suitable for presentation at MICCAI.
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 reviewers generally agreed that the paper addresses an important problem and introduces a novel framework that jointly models spatial alignment and expression integration through a shared Canonical Expression Field. Although concerns were initially raised regarding the consistency of performance gains, component attribution, biological validation, and reproducibility, the rebuttal provided satisfactory clarifications grounded in the submitted manuscript, leading both initially borderline reviewers to raise their final recommendations to Accept. Given the methodological novelty, strong overall experimental performance, and positive post-rebuttal consensus, I recommend acceptance.