acceptodds
Under review as a conference paper at ICLR 2027

Rotation-invariant graph transformers enable global bioimage registration and cell matching

Abstract

Cell matching is a key step in the joint analysis of different imaging modalities in biology. Its main challenges stem from large differences in sample orientation and geometry, local deformations, missing objects, and variations in image appearance and resolution. Commonly used registration pipelines therefore rely heavily on manual or heuristic pre-alignment. We propose grreg, a GNN-based approach for learning context-aware cellular representations that enable direct matching across samples. It combines a rotation- and translation-invariant GCN for local feature extraction and a graph transformer with two kinds of attention: distance-aware self-attention to improve the receptive field of local descriptors, and cross-attention to enable cross-image communication. Our method reaches high matching accuracy, does not require pre-alignment and scales to samples with over 30000 objects through a sparse attention mechanism, as demonstrated on zebrafish embryo development trajectories from Zebrafish Hub. We also introduce a new cross-modal benchmark Platy5+1 to aid further development of bioimage registration.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
biological image registration, graph matching, global registration, geometric learning

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_U49zOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes grreg, a graph neural network-based method for cell matching designed to learn context-sensitive cell representations, thereby enabling direct matching across samples. The method combines a rotation- and translation-invariant Graph Convolutional Network (GCN) for local feature extraction with a graph transformer equipped with two attention mechanisms: distance-aware self-attention to expand the receptive field of local descriptors, and cross-attention to facilitate information interaction across images. Experiments cover HeLa, zebrafish, and Platynereis datasets, and the paper proposes the "Platy5+1" benchmark, comprising five confocal volumes and one electron microscopy volume. While the work demonstrates the capability for large-scale geometric cell matching without pre-alignment, the current version provides insufficient evidence regarding methodological innovation, cross-modal generalization, and precise cell correspondence.

Critical strengths

1.The scientific problem addressed in this paper is significant and well-motivated. Establishing cell correspondences across different individuals and imaging conditions represents a practical bottleneck in the construction of biological atlases; thus, eliminating the need for manual pre-alignment offers clear value. 2.The authors validate their approach using data from HeLa cells, zebrafish developmental sequences, and Platynereis, while also curating the "Platy5+1" benchmark, which comprises five confocal volumes and one electron microscopy volume.

Critical weaknesses
  1. Figure 1 outlines the general scope of the paper, but the technical trajectory remains difficult to discern.
  2. The current paper lacks a discussion on efficiency and the number of parameters; these should be included.
  3. The ablation study presented in Table 1 fails to clearly demonstrate the contributions of individual modules or the overall novelty of the proposed method. The granularity of the ablation study is insufficient.
  4. There is a lack of visualization for the registration results (such as correspondence plots).
  5. The quantitative evaluation is inadequate; the paper fails to include comparisons with a sufficient number of competitive, state-of-the-art methods from recent years, weakening the substantiation of the proposed method's advantages.
  6. The discussion regarding the paper's novelty is insufficient—specifically, how does the proposed method distinguish itself from existing approaches, and what specific improvements does it offer over them?
Questions for authors
  1. The boundaries of the proposed deformable registration method are unclear; specifically, it is uncertain whether the approach remains effective when the object’s appearance undergoes significant changes (e.g., variations in scale, rotation, geometric transformation, or stochastic texture alterations). Figure 2 is poorly structured, making it difficult to discern the logical flow of the paper; it should be revised to improve readability.
  2. The "4.1" section is inadequately described, hindering a clear understanding of the paper's logical structure. In Table 1, both "grreg all-to-all" and "grreg sparse" yield similarly high-performing experimental results. Given that the sparse model can outperform the "all-to-all" (dense) model, please justify the superiority and necessity of the "grreg all-to-all" approach.
  3. What do "GCN 6 conv" and "CPD pre-aligned" in Table 1 refer to? I could not find any definitions for these terms. The paper contains several other concepts that lack explanation, making it difficult to grasp key ideas.
  4. The experimental section mentions three datasets: HeLa, zebrafish, and platy. Why are the results for all three datasets not presented consistently across Tables 1 and 2? Instead, the paper tests different models on different datasets. The completeness of the experimental evaluation should be improved.
  5. The most critical issue in the final part is the lack of comparison with recent SOTA methods; the absence of such comparative experiments at this stage makes it difficult to demonstrate the superiority of the proposed approach.
Reviewer_bBoSOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents grreg, a cell registration model composed of a GNN and graph transformer hybrid with rotation invariant features. The authors propose to alternate self and cross-attention layers to align the representations of the two point clouds. The model is trained with a contrastive loss. The authors also introduce a new benchmark dataset for cross-modal registration and object matching.

Critical strengths
  1. The combination of GNN and Graph Transformer is reasonable for the problem of cell matching, which is related to the problem of graph matching. The problem of cell matching itself is challenging and impactful.
  2. The PCA plots suggest that grreg's embeddings more closely align the objects in the image than GCN.
  3. The authors create a new benchmark dataset (Platy5+1) for cross-modal registration.
Critical weaknesses
  1. The contributions are not clearly stated. The point pair features and the N-tuple contrastive loss come from previous work. The main contribution appears to be applying a GNN+GT to the training setup of (Deng et. al.,2018). The second contribution is the novel benchmark dataset and the experimental verifications. If this is an accurate depiction, the novelty seems limited.
  2. The problem objective in lines 175-179 is too informal. It is not clear what is meant by →min\rightarrow min and →max\rightarrow max. Stating the contrastive loss explicitly would make the exposition clearer. Furthermore, there are a few notational inconsistencies. For instance, X\mathcal{X} is not defined (I assume it is used interchangeably with X\mathbf{X}) , dropping the boldface notation for matrix and vectors (lines 190,216).
  3. The experimental results appear to be limited. In Table 2, it seems grreg underperforms in both acc and 3-hop acc for all cases. Additionally, there is only a single baseline in the experiments (CPD). Given that the features and contrastive training are borrowed from PPFNet, it seems reasonable to also include it as a baseline.
  4. The distance matrix D=DGWd\mathbf{D} = \mathbf{D}_\mathcal{G}\mathbf{W}_d implies that the learnable matrix Wd\mathbf{W}_d is of dimensions N×NN \times N. This would mean that grreg depends on the size of the graph. Therefore the model would need padding for smaller graphs, and would not scale to graphs larger than training graphs. This may be a notation mistake. Could the authors clarify?
  5. The presentation could be improved in a few places. Some parts of the paper are difficult to follow, such as the paragraph explaining the volume feature. Furthermore, the figure captions are not very descriptive, and some seem like placeholders (e.g. Figure 2 only says "grreg architecture"). This makes it difficult for the reader to interpret them. The graph in Figure 2 is a little crowded, which makes it difficult to read.
Questions for authors
  1. The methodology mentions that the K-means clusters are recomputed in every training step. This seems expensive, could the authors expand on the additional computational cost. Is there evidence that this is necessary?
  2. What is the motivation for using a max aggregation operator in the GNN layers (as opposed to sum or mean pooling)?
  3. Could the authors explain the underperformance of grreg with respect to CPD in Table 2?
  4. Multimodality is mentioned in the introduction, yet the experiments mention that in platy "only confocal volumes were used". Does this mean the model is not using the electron microscopy volumes?
1 comment
  1. PPFNet is not applicable to this problem because it assumes that the point cloud has normals, heavily downsamples the point clouds and only addresses the rigid transformations.
  2. We changed the text to include this comment. 3.We chose CPD as a baseline because it is a well-known method widely used by the bioimage analysis community. Our goal in this work was mainly to address the need for pre-alignment by solving the global matching problem directly, therefore we could not identify any methods that compare directly. It is true that it is possible to create pipelines made of various tools but they would have to be modified quite heavily and leave a wide gap for interpretation due to the interaction between the pre-alignment quality and the final matching quality. We would appreciate more specific suggestions concerning the specific methods that would make the baseline more convincing.
  3. Q×KTQ \times K^T has shape N×N×nheadsN \times N \times n_heads. DD has shape N×N×1N \times N \times 1, therefore WdW_d has shape 1×nheads1 \times n_heads and does not depend on the size of the graph. This transformation allows different heads of the transformer to have different weighting between the physical distance and the distance in the embedding space.
  4. We restructured Figure 1 and Figure 2 to better explain the technical pipeline: Figure 1 describes the steps of the algorithm and Figure 2 is focused on the geometric features, as it is the least intuitive part of the method. We also added more detailed figure captions.
Reviewer_cKTTOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper presents grreg, a graph-based model that combines rotation-invariant geometric features with self- and cross-attention to match cells without pre-alignment. Experiments on HeLa, Zebrafish, and confocal samples from Platy5+1 show reasonable registration results, including robustness to large rotations and support for large point clouds.

Critical strengths
  1. The problem is practical. Using cell geometry is a sensible way to reduce dependence on image appearance across modalities.

  2. The method is robust to changes in the initial orientation of the samples.

  3. The self-supervised training setup reduces the need for manually labeled correspondences. Platy5+1 is also a useful contribution, bringing together multiple samples, imaging modalities, and expert annotations.

Critical weaknesses
  1. The cross-modal claim needs more detail and direct experimental support. Although Platy5+1 includes an EM volume, the reported experiments use only confocal volumes, so confocal-to-EM results would help support this claim.

  2. The baseline comparison is too limited, with CPD as the main external baseline. Comparisons with learning-based matching methods and automatic pre-alignment followed by CPD would make the advantages clearer.

  3. The ablation studies do not sufficiently test the main design choices. Separate experiments on the distance bias, cross-attention, oriented volume features, and elastic augmentation would help explain where the gains come from.

  4. The reported inference time of about 18 seconds on the 30k dataset is useful, but the comparison of inference cost is incomplete. Please clarify whether this includes graph construction, clustering, and final matching, and compare runtime and peak memory with baselines on the same hardware.

  5. The training, validation, and test splits are unclear. Please specify which samples and frames are used in each split, and whether any test samples are also used to generate training augmentations.

  6. The cross-attention equation appears to require values from X2WVX^2W_V rather than X1WVX^1W_V, especially when the two graphs have different numbers of nodes. Figure 6, containing “PCA 6”, should be clarified.

Questions for authors

Please refer to "Critical Weaknesses"

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.