List of Papers Browse by Subject Areas Author List
Abstract
Predicting spatial gene expression from histopathology images has the potential to complement Spatial Transcriptomics (ST) while avoiding the costs and limited throughput associated with ST. Multiple recent methods rely on increasingly complex Deep Learning (DL) models. We hypothesized that high complexity is not essential for obtaining competitive performance for mapping histopathology spots to their corresponding gene expression. To test this hypothesis, we compared the image-to-gene expression prediction performance of several benchmarks on convolutional and transformer-based architectures on Visium and Xenium ST datasets. We did not find one clear winner across all settings. We tested several selected lightweight variants, termed as TinyH2GNet, from one of the benchmarks and show that they achieve comparable performance while using only a small fraction of the original parameters. The TinyH2GNet model size is < 1 MB, its inference on a standard CPU is ten times faster than the smallest existing model. This makes TinyH2GNet suitable for efficient image-to-ST prediction on standard CPU hardware without requiring dedicated GPU resources. The results show that increasing model complexity yields diminishing returns and future research should prioritize biological fidelity over architectural scale. All the code for our experiments can be accessed via Github: https://github.com/AI4Med-Lab/TinyH2GNet.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5521_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/AI4Med-Lab/TinyH2GNet
Link to the Dataset(s)
N/A
BibTex
@InProceedings{KauMan_TinyH2GNet_MICCAI2026,
author = { Kaur, Maninder AND Kumar, Amit AND Rani, Ankita AND Goyal, Puneet AND Mall, Raghvendra AND Gupta, Sukrit},
title = { { TinyH2GNet: Exploring Lightweight Models for Gene Expression Prediction from H&E Images } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16891},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper studies the relationship between model complexity and performance for predicting spatial gene expression from H&E images. The authors benchmark several existing CNN- and transformer-based models across three spatial transcriptomics datasets, and then propose lightweight variants (TinyH2GNet) based on EfficientNet compound scaling. The main claim is that reduce models can achieve comparable predictive performance to larger models while offering substantially lower CPU inference time and memory requirements, making them more practical for computational constrained settings
- 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 addresses a relevant question. Many papers focus mainly on architectural complexity, so it is useful to ask whether larger and more complicated models actually provide meaningful gains over simpler baselines. That question is worthwhile and practically important A clear strength is the comparative benchmarking under a common training setup. The authors evaluate multiple models on three datasets and try to keep the protocol consistent, which makes the comparison informative Another strength is the focus on efficiency and deployment constraints. Reporting CPU inference time and model size is useful, especially in a field where practical usability is often under-discussed. The runtime results support the claim that the lightweight models are substantially faster The main empirical observation is also interesting: across the tested datasets, smaller EfficientNet-based models appear to retain similar performance to larger backbones in several settings, suggesting diminishing returns from scale. This is a useful message for the community even if the proposed model itself is simple Finally, code is available which is a positive point for reproducibility
- 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 main weakness is the limited methodological novelty. TinyH2GNet is essentially a pruned EfficientNet-style regression model using standard compound scaling, rather than a genuinely new modelling framework. The main contribution is therefore more in benchmarking and efficiency analysis than in method development, which weakens the paper A second concern is that the evidence for broad generalisation is still limited. The paper uses three datasets, but two are breast cancer Visium-style datasets with relatively small numbers of samples, and the Xenium experiment comes from a single matched H&E image and one sample with many cells. This provides some diversity in platform, but not a very strong test of biological or clinical generalisation across diseases, sites, or institutions Furthermore, although the comparison is useful, it is not always clear that all models are equally well adapted to all datasets and settings. For example, some methods are naturally more suited to certain ST resolutions or contexts than others, and one model could not be applied to Xenium due to memory cost. This is understandable, but it complicates the strength of the overall claim that complexity does not help On the other hand, the clinical deployment framing is also somewhat overstated. Faster CPU inference and small model size are valuable, but this alone does not make the method clinically deployable. The underlying task of predicting spatial transcriptomics from histology remains biologically challenging, and the paper itself acknowledges persistent limitations for low-expression genes. In that sense, the practical significance should be stated more cautiouly. The evaluation is focused mainly on aggregate prediction metrics and runtime, but biological relevance is not explored in much depth. For example, there is limited analysis of which genes or gene programs are predicted well versus poorly, whether certain tissue structures benefit more from larger models, or whether compression harms biologically meaningful outputs selectively. More biological analysis would strengthen this conclusion There are also some writing and presentation issues. Several sentences are awkward or grammatically incorrect, for example “The TinyH2GNet weights were ≤ one MB on the…”, “The results that increasing model complexity yields diminishing returns…”, and “while negative values of ϕ pruning the number of layers…”.
- 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
This paper addresses an important practical question and provides a useful efficiency-oriented benchmark for image-based spatial transcriptomics prediction. I think the main value of the work is showing that much smaller models can remain competitive in several settings while dramatically reducing inference cost. However, the paper would be stronger with a more cautious deployment claim, more biological analysis of what is preserved or lost under model compression, and a clearer framing of the contribution as an efficiency/benchmarking study rather than a major methodological advance
- 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?
I view this paper as slightly below the acceptance threshold. The topic is useful and the benchmarking effort is valuable, especially the focus on fair comparison and computational efficiency. The finding that lightweight models can achieve similar performance to much larger models is interesting and relevant for the field. However, the methodological novelty is limited, the biological analysis remains relatively shallow, and the practical/clinical framing is somewhat stronger than what the experiments currently support. In addition, the paper would benefit from better writing and a more polished presentation. Overall, the work contains a useful message, but in its current form I am not yet convinced it is strong enough for acceptance
- 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.
I thank the authors for their rebuttal
I agree that the empirical question addressed by the paper is timely and practically relevant, if increasingly complex image-to-ST models provide proportionate performance gains relative to lightweight alternatives. Further, the benchmark includes several representative CNN, transformer and hybrid models, and the reported efficiency/performance trade-off of the TinyH2GNet variants is useful
However, several key concerns are still not sufficiently resolved in the manuscript: -the completeness of the benchmark remains only partially justified. The authors state in the rebuttal that omitted recent methods belong to already represented architectural families, but the manuscript does not clearly document this rationale or provide a systematic mapping of omitted methods. Given that the main contribution is empirical benchmarking, this explanation should be explicit
- the scope of the conclusions remains too broad. The experiments are performed on three breast datasets, spanning different spatial transcriptomics platforms and resolutions, but not different tissue domains. Therefore, the conclusion that increasing scale and architectural complexity generally yield diminishing returns should be framed more cautiously as applying to the evaluated breast image-to-ST settings
- the clinical/practical framing is still too strong. The manuscript continues to suggest immediate clinical translation in resource-constrained settings, but the experiments do not establish to be for clinical deployment
- the presentation still contains unresolved writing and clarity issues, including grammatical errors, i.e., punning instead of prune and faster over instead faster than
Overall, I find the empirical direction promising, but the rebuttal does not fully resolve the concerns about benchmark completeness, overgeneralisation, and clinical framing
Review #2
- Please describe the contribution of the paper
The authors propose TinyH2GNet, a lightweight CNN-based model for predicting spatial gene expression from H&E images. They conduct a systematic benchmarking study comparing CNN, transformer, and hybrid architectures across multiple datasets, demonstrating that smaller models with approximately 300k parameters can achieve performance comparable to larger models. Additionally, these lightweight models offer significantly faster inference speeds, making them well-suited for deployment in resource-constrained environments.
- 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.This paper addresses an important and timely research question regarding whether increasing model complexity is necessary for accurate spatial gene expression prediction. The motivation is clearly articulated, particularly in the context of practical deployment, where large models may not be feasible in resource-constrained clinical settings. 2.A major strength of the work is the comprehensive empirical benchmarking study. The authors systematically compare multiple architectures (CNN, transformer, and hybrid models) under identical conditions across three datasets—Hist2ST, HER2 breast cancer, and 10x Xenium. The results, as shown in Table 2, support the key claim that increasing model complexity yields diminishing returns, with no single model consistently outperforming others.
3.The proposed TinyH2GNet demonstrates that a lightweight model with only 3–20% of the parameters of larger models can achieve comparable performance. This highlights its efficiency and practical value, making it well-suited for deployment in real-world clinical environments with limited computational resources.
4.The paper provides a thorough analysis of the trade-off between model size, performance, and inference time. For example, Figure 2 illustrates minimal degradation in L1 error and Pearson’s correlation across reduced model sizes, while Table 3 shows substantial improvements in inference speed, reinforcing the advantages of lightweight models.
- 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.Although the paper aims to address deployment challenges in resource-constrained environments, the proposed model is based on EfficientNet with pruning via compound scaling. While this approach is effective, it offers limited methodological novelty and largely represents an application of existing techniques rather than the introduction of a fundamentally new architecture.
2.The primary contribution of the paper is empirical, centered around benchmarking and scaling analysis. As a result, the novelty of TinyH2GNet is relatively limited, since it is essentially a pruned variant of an existing backbone rather than a newly designed model.
3.While the paper provides detailed performance analysis, all experiments are conducted on three datasets without external validation. This raises concerns about the robustness and generalizability of the proposed approach in broader clinical settings.
- 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
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 presents a practical and well-motivated study on model efficiency for spatial transcriptomics prediction. A key strength is the systematic benchmarking, demonstrating that lightweight models can achieve performance comparable to more complex architectures. This finding has important implications for real-world deployment, particularly in clinical settings where computational resources are limited.
However, the methodological novelty is limited, as the proposed approach relies heavily on existing architectures and scaling strategies. The contribution is primarily empirical rather than introducing a fundamentally new method. Additionally, the lack of cross-dataset or external validation raises concerns about the generalizability of the proposed lightweight model.
Despite these limitations, the paper provides valuable insights into model efficiency and challenges the common assumption that larger models are always better. In practice, even if larger models offer marginal gains, they are often difficult to deploy due to hardware constraints in clinical environments. Given the strong motivation, solid experimental analysis, and practical relevance, I recommend a weak accept.
- Reviewer confidence
Very confident (4)
- [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
The authors present a systematic benchmarking of image-to-ST prediction methods to demonstrate that performance in the task does not scale proportionately with model complexity and size. The authors tested several selected lightweight variants with weights less than 1 MB, labelled generally in the paper as TinyH2GNet, to show that it achieves comparable performance using only a fraction of the original parameters. The authors present extensive experimentation to evaluate performance and latency on CPU and GPU systems. This outlines the reality that even resource constrained environments can run image-to-gene prediction models that have competitive performance.
- 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 has two major strengths:
1.Foundational benchmarking: If done correctly, such benchmarking studies can serve as foundational steps in altering research trends. For instance, if larger and more complex models really don’t achieve results that are better in a statistically significant way, then it would be more prudent to focus on the gaps in the knowledge and data than in model architecture. I believe this study furthers that goal and I commend the researchers on their extensive experiments.
2.Platform diversity: The researchers test on classical ST (square grid), Visium v2 (beehive) and Xenium (single cell level ST). The results are consistent in all platforms and indicate that their claims are reasonable regardless of the data procurement technology or the resolution of the data.
- 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.
I believe there are a few major weaknesses in the study.
Conceptual flaw: The fundamental problem with the motivation behind this study is the assumption that clinicians would want to run “comparable” lightweight models in resource constrained environments. Using image-to-ST predictions for precision medicine and clinical setups require very low margins of error and high performance, much higher than is typically observed in existing models that solve this task. So, the motivation that lightweight models are somehow more valuable because they “are not too bad” and can run in low resource environments is a flawed understanding of the field.
Baselines: The authors do not do justice to the image-to-ST prediction task because they fail to include several more recent studies such as TRIPLEX, MERGE, NH22ST, STEM, mcISTExp, etc. They also do not compare to any foundation models. I understand that FMs are fundamentally large and that defeats the very premise of this paper. But to conclusively demonstrate that size and complexity does not proportionately improve performance, the authors must also compare against the performances of the foundation models, including but not limited to UNI, GigaPath, Virchow, OmiCLIP, and now STORM.
Generalization of model architectures: It is naive to claim that the performance gains associated with various architectures can be simply attributed to their size and complexity. Various approaches in the relevant literature employ various techniques and strategies to model the short- and long-range relationships based on local and global tissue structures and use these relationships to enrich gene expression prediction. Needlessly numerous parameters are worth criticism only when they are not well-grounded in biology. Until the authors exhaustively test against several other baselines that have come out more recently, their claim is not supported conclusively by evidence
- 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
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.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The authors make a commendable effort to demonstrate the size and complexity of models do not dictate performance in the image-to-ST prediction task. They make extensive efforts to test across platforms and with various baselines. However, the evidence is incomplete due to the exclusion of more recent methods as well as the foundation model approaches in this task. The authors’ claims are also slightly misled in the assumption that it is of critical importance to deploy comparable lightweight models to low-resource computational environments.
- 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.
1.The authors explicitly claim in their manuscript that “This makes TinyH2GNet deployable in resource-constrained clinical settings.” Their response to my issue with this claim outright denies this and argues that my observations “conflate” their contributions. I do not understand why the authors are denying the exitence of a claim they directly made in their manuscript.
2.I agree with the explanation that the baselines I mentioned are not fully worth comparing against given that they are complex and nuanced versions of the architectures/backbones they already benchmark. So I am okay with these baselines being skipped.
3.The issue about generalization of model architectures is going to be addressed by adjusting the wording. So I consider this issue resolved.
Author Feedback
We thank the reviewers for their positive feedback: extensive experiments on comparative benchmarks (R1, R2); well motivated, optimal performance-efficiency tradeoff and thorough analysis (R2); and platform diversity (R4). We would like to clarify thesepoints below.(1) On novelty (R1, R2, R4). The paper is not proposed as a new architecture. Its contribution is empirical: a systematic test of whether scale and architectural complexity translate into proportionate gains for image-to-ST prediction. R1 and R2 describe the question as “important,” “timely,” and “practically important,” and R4 calls the benchmarking effort “commendable” and “foundational.” The findings stand on the strength of the experiments, not on a novel module. We will revise the introduction to make this framing explicit rather than leave it implicit. (2) On missing baselines (R4): TRIPLEX, MERGE, NH2ST, STEM, and mcISTExp are extensions of the same architectural families (CNN, graph, transformer, hybrid) that we already benchmarked. Our design choice was to span paradigms, not to list every published instance. Adding more models from the same paradigm would not falsify the central observation that scale and complexity within these paradigms do not yield proportionate gains. On foundation models (UNI, GigaPath, Virchow, OmiCLIP, STORM): R4notes that FMs “are fundamentally large and that defeats the very premise of this paper”. A comparison against models that outright violate the resource constraints we study would not answer the question we pose. We will flag FM evaluation as an open follow-up. (3) On causal attribution (R4): We do not claim that parameter count alone explains performance differences. The paper’s claim is bounded: across representative architectural paradigms, increased scale did not produce proportionate gains on the datasets tested. We will tighten this wording in the discussion. (4) On the “conceptual flaw” (R4): R4 argues that clinicians require high accuracy and therefore lightweight models are not useful. This conflates two distinct points. We are not arguing that current image-to-ST models, lightweight or large, are clinically deployable; no current method, including those R4 recommends as baselines, meets that bar. The paper’s argument is that within the present accuracy regime, additional compute spent on larger models does not buy proportionate accuracy. That is a finding about research direction, not a deployment claim, and we will sharpen this distinction in the revision. (5) On datasets (R1, R2): The three datasets span Visium spot-level data (HER2 ~14k spots; Hist2ST ~30k spots) and Xenium single-cell resolution (~160k cells), and R4 explicitly credits the platform diversity of this design. The diminishing-returns pattern held consistently across these different settings. We acknowledge that all tissue is breast and will state this limitation plainly, but the consistency of the trend across resolution and acquisition technology is itself a form of generalisation evidence. (6) On biological analysis (R1): The revision will expand the discussion of which gene categories and spatial patterns are preserved or degraded under compression. This is added context for the existing findings. (7) On writing (R1): The specific issues (“weights were ≤ one MB”, “while negative values of ϕ pruning…”) were flagged. These will be corrected, and we will go through the paper for similar issues. We hope the above clarifies the paper’s claims and scope. The central empirical finding, that scale and architectural complexity within established paradigms do not yield proportionate gains for image-to-ST prediction, is supported by the experiments as presented.
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.
Paper Summary This paper studies whether increasing model complexity improves performance in predicting spatial gene expression from H&E images. To answer this, the authors perform a systematic benchmarking of CNN, transformer, and hybrid architectures across multiple spatial transcriptomics datasets under a unified training setup. Based on the observation that performance gains from larger models are limited, they propose TinyH2GNet, a lightweight EfficientNet-based model obtained via compound scaling, aiming to retain predictive performance while significantly reducing model size, memory usage, and CPU inference time for resource-constrained settings.
Key Strengths The paper addresses an important and practically relevant question regarding the necessity of model complexity in spatial transcriptomics prediction. Reviewers consistently highlight the value of the systematic benchmarking under controlled conditions, which enables meaningful comparisons across architectures and datasets. The empirical finding that lightweight models can achieve comparable performance while significantly improving efficiency is considered insightful and useful for the community. The analysis of trade-offs between performance, model size, and inference time, along with evaluation across multiple platforms, strengthens the practical relevance of the study.
Key Weaknesses The main weakness is the limited methodological novelty, as the proposed model is largely a scaled variant of an existing architecture rather than a new design. There are concerns about the completeness of the empirical evidence, including limited dataset diversity, lack of external validation, and omission of several recent or stronger baselines, which weakens the support for the central claim. Reviewers also question the strength of the conclusion that model complexity does not significantly impact performance, noting that architectural differences and modeling strategies are not fully accounted for. Additionally, the clinical and deployment motivation is considered overstated, and the paper lacks deeper biological analysis to support practical significance. Writing and presentation issues are also noted.
Review Summary All reviewers agree that the paper tackles a relevant problem and that the benchmarking effort and efficiency analysis are valuable. There is consensus that the observation of diminishing returns from model complexity is interesting. However, reviewers also consistently raise concerns about limited novelty and insufficient support for the main claims. Reviewers see that missing baselines, limited validation, and incomplete analysis weaken the conclusions. Overall, concerns about the strength and completeness of evidence remain significant despite agreement on the paper’s motivation and empirical value.
Decision Invite to rebuttal. The paper has useful empirical insights, but concerns about novelty, missing baselines, and strength of claims need clarification.
Rebuttal Guidance The rebuttal should focus on justifying the completeness of the benchmarking, particularly addressing concerns about missing recent methods and explaining how the selected baselines sufficiently support the main claims. The authors should clarify the scope of their conclusions regarding model complexity and avoid overgeneralization, especially in relation to different architectures and datasets. Strengthening the discussion around generalization and practical applicability, including a more cautious framing of clinical relevance, would be important. Additional clarification on limitations, along with improved explanation of experimental design choices and clearer presentation, would help resolve reviewer concerns.
- 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 satisfactorily addressed the main concerns raised during review. The authors clarified that the contribution is primarily an empirical benchmarking and efficiency study rather than a novel architectural advance, justified the selection of representative baselines, and appropriately narrowed the scope of their conclusions. Concerns regarding clinical deployment claims and overgeneralization were acknowledged and addressed through more careful framing. While limitations in novelty and dataset diversity remain, I believe the rebuttal sufficiently resolved the major concerns, and the paper provides valuable empirical insights for the image-to-spatial-transcriptomics community. I therefore recommend acceptance, with the expectation that the final version narrows its claims to the evaluated breast image-to-ST settings and removes overly strong clinical translation language.
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 paper presents TinyH2GNet, an empirical benchmark showing that lightweight EfficientNet-based models can match larger models for H&E-to-spatial-transcriptomics prediction with much lower computational cost. Reviewers found the study timely and practically useful, though concerns remain about novelty, baseline coverage, dataset scope, and clinical claims.
I give significant weight to R4’s shift to accept after the rebuttal and R2’s consistently positive assessment. The rebuttal clarified the paper’s empirical contribution and addressed several presentation and baseline concerns.
Overall, I recommend acceptance. Despite some limitations, the paper provides useful insights into efficiency–performance tradeoffs for spatial transcriptomics prediction.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
After rebuttal, R1 maintains Reject due to unresolved concerns about benchmark completeness, overgeneralized claims, clinical framing, and writing clarity, while R4 changes to Accept but still urges caution on clinical deployability. Given the concerns unresolved, this current version could not be accepted.