acceptodds
Withdrawn from ICLR 2027

Knowledge-Guided Semantic Confusion-Aware and Reliable Spatial Gradient Learning for Cell Abundance Prediction from H&E Images

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Accurate prediction of the abundance and spatial distribution of different cell types within tissues is essential for characterizing tissue heterogeneity and disease microenvironments. Existing two-stage methods are susceptible to error propagation from intermediate gene expression prediction, whereas one-stage methods that directly predict cell abundance from H&E images still suffer from insufficient modeling of cell-type confusion and inadequate preservation of local spatial structures. To address these issues, we propose a Knowledge-Guided Semantic Confusion-Aware and Reliable Spatial Gradient Learning framework for cell abundance prediction from H&E images. Specifically, we develop a knowledge-guided cell semantic confusion-aware relative abundance learning mechanism to reduce abundance confusion among biologically related cell types. We further propose a reliable region-aware spatial gradient consistency constraint to improve the spatial consistency of local cell abundance variations. Experiments on multiple public spatial transcriptomics datasets demonstrate that the proposed method effectively improves fine-grained cell abundance prediction and enhances the recovery of spatial cell distributions.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Cell abundance prediction, H&E images, Semantic confusion-aware learning, Spatial gradient consistency

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_RewjOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

Aiming at the problem of confusion between similar cell types and the lack of constraints on local spatial variations when predicting cell abundance directly from H&E images, this paper proposes two training constraints: KGSC uses marker genes from single-cell references to generate six-field descriptions via GPT-5.5, which are then processed by a frozen PubMedBERT to obtain cell semantics, and selects easily confused cell pairs based on abundance correlation in the training set to supervise their relative abundance differences. RSGC defines enriched regions and boundaries based on label abundance, supervising reliable spatial gradients only on the local GATv2 branch. Experiments on multiple datasets show that the proposed method outperforms existing methods in cell abundance prediction.

Critical strengths

Compared to independent regression for each cell type, the relative differences of easily confused pairs in KGSC and the regional gradients in RSGC have certain biological significance.

Critical weaknesses
  • The related work section analyzes one-stage cell abundance prediction methods HP-Cell and CUCA, but there is no comparison with them in the experiments.
  • The main contribution of the paper is the proposal of two training constraints, but the current ablation experiments cannot rigorously prove the effectiveness of these two constraints.
Questions for authors
  1. The ablation experiments seem to be based on single-run results. Have you tried running multiple experiments and averaging the results? If it is a single run, have you considered the impact of randomness on the results?
  2. If the two losses proposed in this paper are added to the backbone of Hist2Cell, would similar improvements be obtained? In other words, does the improvement in this paper depend on a specific backbone structure?
  3. The semantic descriptions are generated by GPT-5.5. Is Table 3 the complete prompt? The generated description text is not disclosed, nor is the code. If the model output is changed, could SsemS_{\mathrm{sem}} and the set of easily confused pairs change? Would this affect the final prediction?
  4. Were the slices partitioned by donor? If not, please explain whether there is a potential data leakage issue.
  5. What is the rationale for equally averaging the features and predictions in formulas 1 and 2? Could summing them dilute the contribution of certain features or predictions?
  6. It is recommended to add some demo figures to illustrate the process and effects of KGSC and RSGC.
  7. How is the semantic-confusion pair defined?
Reviewer_Pj8ZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The authors proposed a one stage knowledge-guided semantic confusion-aware and reliable spatial gradient learning method for fine-grained cell abundance prediction from H&E images. Their contribution was adding two regularizers in existing one stage model of cell abundance prediction from H&E images:

  1. KGSC: they used LLM-derived semantic similarity of scRNA-seq marker descriptions, multiplied by abundance correlation to supervise the predicted and ground truth difference between their standardized abundances at each spot. its applied in the final part.
  2. RSGC: its applied in the local part. It supervises the predicted difference between standardized abundances of neighboring spots to match the ground-truth spatial gradients

They performed experiments in 3 different datasets, measuring the overalln supervised perfomance.

Critical strengths
  1. The authors target two well defined limitations of previous methodlogies and motivate their model decisions clearly. Additionally, both of the proposed regularizers are modular and can be added in other baseline supervised models.
  2. The methodological design respects the fact that cell types differ in abundance and spatial pattern and they use cell type specific standardization in the formulation
  3. The authors used 3 different datasets for training (3 different models) with each containing different tissues and pathologies.
Critical weaknesses
  1. Lack of benchmarking of individual compoments such as ResNets, Transformer and the LLM. Since the comparison of different models is missing, an explanation of why each model was chosen would help understand each choice.
  2. One of the main weaknesses in my opinion is only using PCC and SCC as evaluation metrics. Both of these are averaged out across many values and don't capture all the information needed to be evaluated. None of these metrics measure what the 2 additional regularizers offer. Therefore improved performance being atributed to these two addition is hypothetical. The author offer ablation studies showing that their addition imporves the performance, but it misses pair-level metrics for confusion and spatial or boundary metrics (for RSGC). Qualitative evaluation helps towards that but 4 hand selected cases is not enough and the two experiments where individual components are evaluated are helping but were not performed systematically.
  3. Hist2Cell is very close both in the quantitative metrics and the visualizations. I want to note that in table 1 and 2 the Hist2Cell is better than table 2 's base model. The authors give little to no justification of their base model selection (before adding the 2 regularizations). This becomes even more confusing in section 4.5 and 4.6 where they compare the KGSC-only and RSGC-only versions with the Hist2Cell instead of the base model with no justification of this decision. Is the base model same with Hist2Cell? If so why Hist2Cell is better than the base model's. The improvement of their model compared to Hist2Cell is questonable since its within error in many cases.
Questions for authors
  1. In regards to the design:
  • why the authors chose to clip the correlation to 0 in KGSC? can they explain their choice? what about the negative correlated pairs?
  • also can the authors justify the usage of PubmedBert instead of a another semantic similarity metric? I would lilke to see a ablation replacing the LLM with a simpler mathematical metric (for example cosine similarity of the reference expression profiles directly) to showcase the value the LLM adds.
  • is the top-k the right way? what if some have many (>k) correlated embeddings and other have none or < k? Maybe a threshold would be a more sound solution here. Can the authors justify this decision and show examples of correlations, cases where a spot has many pairs and a example where a spot has very little?
  • Can the authors clarify which one is the Z(.) funtction in the KGSC part? Is the training-set statistics z-score or its some other function in the model? And are the training-set statistics performed globally or per cell type?
  • Can the authors clarify what V_center is or reference its definition in the appendix?
  • Similarly could the authors specify how s_c is computed or reference in the appendix?
  1. Why table 1 - ours and table 2 last low doesnt match? where the ablation only perfomred for one seed (the best one)? It would be helpful to also perform across multiple seeds to see if the difference in scores are within error. One thing that needs explanation in that the base model without either module in table 2 gets HER2ST PCC 0.552, which is higher than what the full model reported in table 1 (0.551)

  2. The authors should mention how their base model is different than Hist2Cell and if there are the same state it specifically and note that the only differences are the two regularizers. Also clarify the inconsistencies from weakness 3. If they are the same Table 2’s base model should then equal Hist2Cell in Table 1, but they are different: Lung PCC 0.326 vs 0.334, HER2ST 0.552 vs 0.542, STNet 0.268 vs 0.260. These are within seed noise. Which numbers come from which runs?

Overall if any results that make the authors' argument of the need of these two regularizers stronger is in the appendix due to space issue, I would suggest to be referenced in the main text.

Reviewer_qeZLOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes a framework for predicting fine-grained cell-type abundance from H&E images. It extends a Hist2Cell-style multi-level architecture with two training objectives: KGSC constructs potentially confusable cell-type pairs from semantic representations and training-set abundance correlations, then supervises their relative abundances; RSGC applies region- and gradient-dependent consistency supervision to the local prediction branch. The authors evaluate the method on Human Lung, HER2ST, and STNet, report PCC and SCC over three random seeds, and provide ablations and additional spatial analyses. The targets are Cell2location-derived abundance estimates rather than independently measured cell counts, which is relevant to how the results should be interpreted.

Critical strengths

1.The paper addresses a useful problem at the intersection of computational pathology and spatial biology: predicting fine-grained cellular composition directly from widely available H&E images. The focus on cell-type confusion and spatial organization is well motivated for this task. 2.The proposed auxiliary objectives are clearly tied to the stated failure modes, and the paper includes ablations and additional analyses of semantic-pair and spatial-gradient behavior. The use of semantic information only during training also keeps the stated inference inputs relatively simple.

Critical weaknesses

1.Test-set-based checkpoint selection. The supplementary training scripts for both the proposed method and Hist2Cell explicitly select the checkpoint with the highest test-set PCC across epochs. I could not identify an independent validation-set selection procedure in the supplied scripts or the experimental description. If this protocol was used to generate the reported results, the test set participated in model selection, undermining the interpretation of the results as independent estimates of generalization. Applying the same procedure to both methods does not resolve this concern. 2.Insufficient specification of the evaluation and aggregation protocol. The connection between the supplied splits, training scripts, and reported tables is not sufficiently clear. For Human Lung, the supplied training script uses the A50 split, while additional leave-donor-out split files are included without a clear explanation of whether they contribute to the reported results. The paper also does not clearly specify whether the main PCC/SCC values are computed after pooling spots across test sections or separately within each section and then averaged. These choices affect the interpretation of the reported improvements.

Questions for authors

1.Checkpoint selection and validation: Were the supplied training scripts, or code using the same test-PCC-based checkpoint-selection rule, used to generate Tables 1 and 2? Was an independent validation set used for selecting epochs and hyperparameters? If so, please specify its composition and provide the corresponding selection procedure. If test PCC was used for selection, can you provide results obtained with validation-based model selection and an untouched test set? 2.Splits and result provenance: Which exact train/test splits were used for each dataset and table? For Human Lung, were the reported results obtained only on the A50 split or aggregated across multiple held-out donors? Please clarify the role of the other supplied split files and whether all methods used identical splits. 3.Metric aggregation and uncertainty: Were PCC and SCC computed by pooling spots across test sections, or computed separately for each section and then averaged? How were cell types with constant targets or undefined correlations handled? Please provide per-seed and per-test-section results so that the consistency of the reported gains can be assessed.

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

Related papers

Discussion (0)

Sign in to comment.