List of Papers Browse by Subject Areas Author List
Abstract
Spatial tissue image analysis is essential for quantifying cellular microenvironments and tissue architecture. Yet existing software typically exposes a fixed feature set, which may limit biomarker discovery. When required features are missing, bioscientists may require collaboration with computational experts, which can introduce delays and limit scalability. To address these limitations, we propose \textbf{TissueCodePilot}, a coding-and-reasoning agent that observes relevant information from the analysis environment and autonomously plans and executes actions to support flexible, on-demand spatial tissue analysis. Unlike prior coding agents that rely on detailed, iterative prompting, TissueCodePilot requires only a single minimal user prompt, making it accessible to bioscientists without extensive programming experience. To evaluate TissueCodePilot, we curate an expert-annotated dataset spanning three tissue types with multiple fields of view per tissue. Each field of view is paired with 50 questions that provide minimal prompt context and cover diverse spatial feature categories. In total, the dataset comprises 1{,}500 image–question pairs with corresponding ground-truth outputs. We will publicly release this benchmark to support future research as the first coding-agent benchmark for spatial tissue image analysis. On this new benchmark, TissueCodePilot substantially outperforms prompt-instruction coding agent baselines, boosting Success Rate from $\sim$0-2.4\% to 18.5-35.4\%, increasing pass@5 from at most 11.61\% to 54.82-75.05\%, and increasing pass@10 from at most 22.48\% to 64.00–86.00\%. For reproducibility, our code will be made publicly available at \url{https://github.com/hula-ai/TissueCodePilot}.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1741_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/hula-ai/TissueCodePilot
Link to the Dataset(s)
N/A
BibTex
@InProceedings{VoHun_TissueCodePilot_MICCAI2026,
author = { Vo, Hung Q. AND Vo, Huy Q. AND Zhao, Hong AND Wong, Stephen T. C. AND Nguyen, Hien V.},
title = { { TissueCodePilot: A Code-Action Agent for AI-Assisted Spatial Tissue Analysis } },
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
The paper presents an interactive coding agent for spatial tissue analysis that can use tools, execute code, and iteratively refine its solution from a minimal natural language query. It also introduces a new benchmark for this setting and shows strong gains over prompt based coding baselines.
- 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.
-Clear practical motivation and relevant problem setting. The paper addresses an important bottleneck in spatial tissue analysis, where users may need flexible feature extraction beyond fixed software pipelines. -Strong empirical gains across all reported baselines. TissueCodePilot shows large and consistent improvements across three tissue types and multiple LLM backbones, which makes the empirical findings convincing. -Useful benchmark contribution. The paper introduces a new benchmark for coding-based spatial tissue analysis, which is valuable and can support future work in this area.
- 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.
-Evaluation is somewhat constructed. Queries are not real bioscientist prompts. Authors first define spatial features and then use an LLM to convert them into short questions. So real-world generalization to messy, open-ended user requests remains unclear.
-Contribution is more system integration than methodological novelty. The agent mainly combines existing ideas like ReAct-style reasoning, tool use, Python execution, and self-correction. Useful system, but the core method feels incremental.
-Baseline setup is unbalanced. TissueCodePilot is interactive, while baselines are more limited. So the gains are not cleanly isolated. -Evaluation protocol is unclear, especially how pass@k is computed for one-shot baselines. Hard to separate gains from true interaction versus repeated stochastic sampling across runs.
-Success Rate is not explicitly defined in the manuscript. Minor: Some statements in the introduction are repetitive. “In contrast, human experts iterate writing, running, debugging, and adapting step by step. “ etc
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper addresses a practical and relevant problem, introduces a useful benchmark, and shows strong empirical gains over the reported baselines. The score is reduced by the fact that the method feels more like system integration than clear methodological innovation, the evaluation setup is somewhat constructed, and parts of the evaluation protocol are not described clearly enough.
- 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
Review #2
- Please describe the contribution of the paper
TissueCodePilot is an interactive coding agent for spatial tissue analysis. It is evaluated on a benchmark of 1,500 CosMx SMI image-question pairs.
- 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.Spatial tissue analysis often requires custom measurements that are difficult to cover with fixed analysis tools. Exploring the agent is interesting.
2.The benchmark is a useful direction because it tests whether an agent can turn image-based questions into executable analysis steps.
3.The results suggest that code execution and iterative correction are helpful for this type of 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.
1.TissueCodePilot can call tools, run Python code, inspect errors, use intermediate outputs, and re-plan, while most baselines are one-shot or limited-retry prompt baselines. This is not a matched comparison. The paper needs a generic ReAct or CodeAct agent with the same tools, the same Python environment, and the same step budget.
2.Observation, tool use, Python execution, and replanning are already standard components of agentic coding systems. The paper does not make clear what is methodologically specific to TissueCodePilot beyond applying this workflow to spatial tissue analysis.
3.The best single-run success rate is only about 18.5% to 35.4%. Pass@5 and pass@10 are higher. I would like to know how a real user would identify the correct answer among multiple sampled outputs.
4.The benchmark definition is incomplete. The paper should provide the details about the feature taxonomy, ground-truth scripts, numerical tolerance, units, or answer-matching rules.
5.It is also not clear how segmentation or cell-type errors propagate into the final answers. Will these errors be corrected or dropped?
6.Future work and the limitations of this work should be discussed.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The benchmark direction is practical, but the method contribution is not yet clear. The main result compares an interactive coding agent against much weaker non-interactive baselines. The low single-run success rate and incomplete benchmark details further weaken the evidence.
- 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
Review #3
- Please describe the contribution of the paper
This paper introduces TissueCodePilot, an LLM-based agent for building and executing computational histopathology analysis tasks from natural language instructions. The paper also introduces a benchmark for evaluation of similar agents.
- 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.
- To my knowledge, this is a first-of-its-kind assistant in the literature.
- The release of code and benchmark will likely be a useful tool for developing future tools in the field.
- The quantitative results seem to convincingly demonstrate the utility of the author’s design choices.
- The paper is clearly written and easy to follow.
- 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.
- (Methods) It is unclear how the tool interacts with intermediate artefacts (such as segmentation maps) if it is only a language-only model. What happens if the segmentation quality is very poor - how does the agent detect and rectify such a situation?
- (Methods) A fuller description of the tools available to the agent (e.g. as a table) and the way the agent interfaces with them (e.g. through an MCP server) could improve the reproducibility of the methods.
- (Experiment Results and Discussion) For completeness (and accessibility to a MICCAI audience), I think it would be good for the definitions of the metrics used to be explicitly defined in the paper.
- (Experiment Results and Discussion) The qualitiative results section would massively benefit from feedback from the bioscientists giving the queries (e.g. on ease of use).
- 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?
This is a strong paper, though slightly let down from a lack of clarity in the description of the agent itself - I could try to piece together an implementation from the description, but wouldn’t be confident that I’ve properly recapitulated the design exactly. I imagine this ambiguity would be resolved in the final paper, however, when the code and dataset are released (as promised by the authors).
- Reviewer confidence
Somewhat confident (2)
- [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 for their insightful and constructive feedback and truly appreciate their time spent reviewing this work. Below, we address the main concerns shared across reviews. 1.On novelty and contribution (R1, R2, Meta-Reviewer): We agree that the individual components (e.g., ReAct/CodeAct-style reasoning, tool use, and code execution) are not novel by themselves. Our main contribution lies in system-level integration for spatial tissue analysis, including a new problem formulation, a dedicated benchmark, and empirical analysis of agent behavior in this domain. 2.On fairness of comparisons and baselines (R1, R2, Meta-Reviewer): We acknowledge concerns that TissueCodePilot has stronger agentic capabilities than some baselines. Our method follows a CodeAct-style framework with tool use and execution, and we evaluate coding-agent baselines under a controlled one-shot minimal-prompt setting. We did not include generic ReAct as it relies on text-based reasoning without executable code, which is not aligned with our focus on coding-based assistants. 3.On evaluation protocol and metric clarity (R1, R2, R3): We will improve clarity by explicitly defining all evaluation metrics, including success rate and pass@k. Specifically, pass@k is computed consistently by running each agent k times per query and measuring the proportion of successful runs. 4.On agent capabilities and interaction with intermediate artifacts (R3, Meta-Reviewer): TissueCodePilot is not a language-only model; it generates and executes code, and interacts with intermediate artifacts (e.g., segmentation maps) as program variables within the execution environment. 5.On robustness to cell segmentation and cell-type classification errors (R2, R3, Meta-Reviewer): While segmentation and classification quality affect final outputs, the agent’s coding and reasoning process is independent of these components. These tools are modular; improved upstream performance leads to better results but does not alter the agent’s decision process. 6.On benchmark realism and generalizability (R1, Meta-Reviewer): We acknowledge that our queries are curated rather than directly sourced from bioscientists. To mitigate this, we designed prompts to be simple, natural, and free of coding instructions, and had them reviewed by a bioscientist expert to ensure domain relevance. 7.On usability and practical deployment (R2, R3, Meta-Reviewer): We agree that understanding how users identify correct outputs is important. In this work, pass@k is used solely for evaluation and not intended for user-facing selection.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- 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.
This paper introduces TissueCodePilot, an LLM-based agent for building and executing computational histopathology analysis tasks from natural language prompts. The reviewers pointed out major strengths, especially given that the paper addresses an important challenge in computational histopathology, where users may need flexible feature extraction beyond fixed software pipelines. The use of LLM-based agents is interesting and well-motivated in the paper. The proposed tool is benchmarked against several state-of-the-art methods, and the results convincingly demonstrate its utility. However, several weaknesses have been noted by reviewers, and some could be easily addressed in the final version:
- Clarify the main contribution of the paper. The use of an AI agent is not novel, and the proposed framework offers no new ideas in that field. Therefore, I advise the authors to reframe their claims and place greater emphasis on the practical contributions: system integration for computational histopathology addresses a bottleneck in the field, etc.
- Provide a more detailed description of the inputs available to the agent and the way the agent interfaces with them to facilitate understanding and improve the reproducibility of the methods.
- Clarify the way the agent deals with segmentation or cell-type errors. Reviewers were concerned that, because the tool is solely based on LLMs, it might fail to detect low-quality or errors in intermediate outputs.
- Not sure if this could be addressed in the rebuttal, but reviewers mentioned several issues regarding the comparison with existing methods and results analysis. In particular, R1 and R2 pointed out the unfairness of comparisons, as competitors are mostly one-shot or limited-retry prompt baselines. Comparison with a similar ReAct or CodeAct agent would strengthen the claims of outperformance.
- The reviewers pointed out the limited generalizability of the evaluation: queries are not made by bioscientists, and the tools’ behavior under real-world user requests remains unclear. Moreover, it would be interesting to include a user study to better understand how bioscientists would interact with the tool, both in terms of input and output.