Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Delineating the clinical target volume (CTV) in radiotherapy involves complex margins constrained by tumor location and anatomical barriers. While deep learning models automate this process, their rigid reliance on expert-annotated data requires costly retraining whenever clinical guidelines update. To overcome this limitation, we introduce OncoAgent, a novel guideline-aware AI agent framework that seamlessly converts textual clinical guidelines into three-dimensional target contours without any target volume annotation. Evaluated on esophageal cancer cases, the agent achieves a Dice similarity coefficient of 0.842 for the CTV and 0.880 for the planning target volume, demonstrating performance highly comparable to a fully supervised nnU-Net baseline. Notably, in a blinded clinical evaluation, physicians strongly preferred OncoAgent over the supervised baseline, rating it higher in guideline compliance, modification effort, and clinical acceptability. Furthermore, without any retraining, the framework generalizes to alternative esophageal guidelines and shows preliminary extensibility to other anatomical sites (e.g., prostate). Beyond mere volumetric overlap, our agent-based paradigm offers near-instantaneous adaptability to alternative guidelines, providing a scalable and transparent pathway toward interpretability in radiotherapy treatment planning.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5049_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/Oncosoft-Research/OncoAgent

Link to the Dataset(s)

N/A

BibTex

@InProceedings{KimYoo_AGuidelineAware_MICCAI2026,
        author = { Kim, Yoon Jo AND Cho, Wonyoung AND Lee, Jongmin AND Chae, Han Joo AND Park, Hyunki AND Seo, Sang Hoon AND Noh, Jae Myung AND Yang, Kyungmi AND Oh, Dongryul AND Kim, Jin Sung},
        title = { { A Guideline-Aware AI Agent for Zero-Shot Target Volume Auto-Delineation } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16879},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The main contribution of this paper is an LLM-based agentic framework for Clinical Target Volume (CTV) auto-delineation, validated using esophageal cancer cases. The framework translates free-text clinical guidelines into a structured delineation plan, subsequently, into 3D target volumes. To achieve this, it orchestrates the execution of pre-trained models for Organs at Risk (OARs) segmentation alongside geometric operation tools. Using this approach, only free-text clinical guidelines and specific set of parameters, such as body region or dose level, are required as input. Crucially, if clinical guidelines change, the solution does not require costly retraining (unlike traditional deep-learning segmentation methods) but only an update to the guideline descriptions within the prompt. The paper demonstrates the framework adaptability to other protocols and anatomical sites (prostate) in a zero-shot manner. On the esophageal dataset, the framework achieved results comparable to trained nnUNet-based baseline model.

  • 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.Novel Application: The paper introduces novel application of Large Language Model (LLM) and an agentic framework for Clinical Target Volume (CTV) delineation. A key advantage of this approach is that it bypasses the traditional labor-intensive cycle of data re-annotation and model retraining typically required whenever clinical guidelines are changed. 2.Clinically Aligned Evaluation: The evaluation is strengthened by the inclusion of expert human assessment, which provides deeper insight than sole standard automated metrics. The findings reveal that while the framework achieves quantitative results comparable to state-of-the art deep learning method, its outputs are much better aligned with human perception. 3.Safety mechanisms: The framework incorporates a dedicated safety check to ensure the structural validity of generated plans, achieved by implementation of self-refinement mechanism if violations of plan are detected. 4.Scientific transparency: The authors provide a candid discussion of the study’s limitations. By explicitly addressing factors such as the small evaluation dataset and risk of model hallucinations they show realistic roadmap for future improvements.

  • 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. Insufficient Implementation Detail: While the proposed concept is compelling and clinically relevant, the paper lacks several critical implementation details. Specifically, the authors do not specify essential hyperparameters such as the temperature settings, or reasoning effort for the LLM. Furthermore, the specific architectures and versions of the models utilized for OARs segmentations are not clearly identified, making it difficult to assess the technical baseline. 2.Lack of Stochastic and Reliability Analysis: Although the authors acknowledge the risk of model hallucinations the paper provides no quantitative data regarding their frequency. Additionally, there is a lack of information concerning output stability (consistency) when the model is queried multiple times for the same case. Statistics on how often the self-refinement mechanism was triggered are also missing. Such data is important for establishing clinical trust. 3.Barriers to Reproducibility: The absence of a public code repository or the disclosure of the detailed prompt templates significantly hiders the reproducibility of the study. Given that LLM-based agentic frameworks are sensitive to specific prompting strategies those details are important to be able to validate and build upon the findings.
  • 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 does not provide sufficient information for 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?

    My recommendation of a Weak Reject is primarily driven by the lack of granular implementation details and the limited analysis regarding system reliability. While the core concept of an LLM-based agentic framework for CTV delineation is both highly relevant and innovative, the current manuscript does not provide sufficient technical depth (e.g., specific hyperparameters and OAR model configurations) or quantitative data on model stability and hallucination rates.

  • 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 authors’ response has successfully addressed my primary concerns, convincing me to upgrade my recommendation from a Weak Reject to an Accept. Specifically, they provided the missing LLM hyperparameters, clarified issues regarding model hallucinations and repeatability, and committed to publicly sharing their prompts and code upon acceptance.



Review #2

  • Please describe the contribution of the paper

    This paper presents OncoAgent, a guideline-aware agentic framework that converts radiotherapy guidelines into executable tool-call sequences and generates CTV/PTV using pre-trained OAR segmentation models and geometric operations. The problem setting is clinically meaningful, especially given that guideline updates can quickly make purely supervised delineation models outdated. The proposed framework is interesting and potentially useful from an interpretability and adaptability perspective. In the reported experiments, the method achieves performance close to a strong supervised baseline on a small esophageal cancer test set, and receives better blinded physician ratings.

    That said, I feel several aspects of the work would benefit from clearer positioning and stronger validation. In its current form, the method appears closer to a guideline-to-execution pipeline that translates textual instructions into a predefined sequence of OAR segmentation, GTV expansion, Boolean exclusion, and post-processing, rather than a new end-to-end target delineation model. Some of the stronger claims, such as “training-free,” “zero-shot auto-delineation,” and “cross-site generalization,” may therefore be somewhat overstated relative to the current evidence.

  • 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 addresses a clinically relevant problem. Frequent guideline changes are indeed challenging for conventional supervised delineation pipelines. 2.The proposed framework has an appealing level of interpretability, since the intermediate execution plan is human-readable and, in principle, reviewable by clinicians. 3.Within the reported experimental setup, the quantitative and qualitative results are promising. OncoAgent performs comparably to nnU-Net(GTV Prior) on CTV/PTV metrics and achieves better physician ratings.

  • 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 task formulation would benefit from more precise positioning. The execution engine operates on the patient CT and existing GTV contours, while relying on pre-trained OAR segmentation models. As such, the method appears closer to guideline-aware target construction given GTV than fully automatic zero-shot target delineation directly from CT. Clarifying this distinction would make the paper easier to interpret. 2.The added value of the LLM/agent component is not yet fully isolated. At present, the framework seems to mainly parameterize textual guidelines into a largely fixed execution template, rather than perform genuinely complex clinical reasoning. The reported average of 1.13 LLM inference calls per case also suggests that the process may be highly templated. Without a non-LLM baseline, such as a manually implemented rule engine or template-based parser using the same geometric pipeline, it is difficult to determine how much of the gain comes from the agentic component itself versus the explicit rule-based construction. 3.It is not yet fully demonstrated that the framework can accommodate the full complexity of esophageal target expansion rules. In practice, esophageal CTV delineation often involves conditional, hierarchical, and context-dependent decisions, and may not always be reducible to simple margin expansion followed by OAR subtraction. The manuscript notes that guideline ranges are resolved using patient-specific context, such as dose level or physician preferences, but this decision mechanism is not described in enough detail to assess its robustness and reproducibility. 4.The validation remains limited in scale. The study includes 40 patients in total, with only 8 test cases, and the blinded physician evaluation is based on those same 8 cases rated by 2 radiation oncologists. For a paper emphasizing clinical preference and clinical competitiveness, this is still a relatively limited validation. 5.The evidence for cross-guideline and cross-site generalization is still indirect. The alternative-guideline and prostate experiments are mainly supported by Tool Call F1, which reflects plan-sequence agreement rather than actual contour quality or downstream clinical utility. In particular, the prostate result of 0.64 suggests that substantial challenges remain when extending the framework to a new anatomical site.

  • 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 provide sufficient information for 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?

    The paper addresses a clinically relevant problem, and the motivation is clear. I think the idea of making target delineation more adaptable to changing guidelines is interesting. The reported results are promising in the presented setting. However, I remain slightly below the acceptance threshold for three main reasons: first, the method would benefit from clearer positioning, since it relies on existing GTV contours and pre-trained OAR segmentation and therefore feels closer to guideline-aware target construction than fully automatic zero-shot delineation; second, the added value of the LLM/agent component is not yet fully isolated from the underlying rule-based geometric pipeline; and third, the current evidence is still limited in scale and in the strength of the cross-guideline/cross-site validation. Overall, I think this is a promising direction, but the current version still needs clearer framing and stronger validation.

  • Reviewer confidence

    Very confident (4)

  • [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.

    My main concern after rebuttal is the insufficient specification of the patient-specific decision mechanism. The authors state that the LLM can resolve guideline ranges and context-dependent rules using patient-specific information, but it remains unclear how GTV location is extracted, how upper/mid/lower esophageal involvement is determined, how margin ranges and physician preferences are operationalized, and whether such decisions are correct at the case level. These are not minor implementation details, because they directly affect the generated CTV/PTV and are central to the claimed guideline-aware reasoning capability.

    The rebuttal provides illustrative examples, such as selecting nodal regions according to tumor location, but it does not provide a sufficiently reproducible mechanism for how these decisions are made from the available inputs. It also remains unclear whether the human-readable tool-call plan is only an audit trail after the LLM decision has been made, or whether there is a systematic validation process to ensure that the upstream clinical reasoning is correct before execution. In this sense, interpretability of the execution plan does not fully address the reliability of the patient-specific decision process.

    I appreciate the authors’ clarification that “zero-shot” means no CTV-specific training and their willingness to tone down overstrong claims. However, the current evidence still does not fully demonstrate that the framework can robustly handle conditional, hierarchical, and context-dependent clinical rules in a reproducible way. Since these decisions directly determine the final contours, this remains a central limitation. I therefore maintain my reject assessment.



Review #3

  • Please describe the contribution of the paper

    The paper proposes OncoAgent, a guideline-aware AI agent that converts textual radiotherapy guidelines into 3D target volume contours in a zero-shot manner. The framework uses a large language model to translate guideline text into structured tool-call sequences, which are executed using pre-trained OAR segmentation models and geometric operations (e.g., dilation and subtraction). Evaluated on esophageal cancer cases, OncoAgent achieves performance comparable to a supervised nnU-Net baseline while demonstrating improved physician preference in a blinded evaluation. The approach further claims zero-shot adaptability to alternative guidelines and anatomical sites.

  • 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 work introduces a fundamentally different approach to target delineation by explicitly leveraging clinical guidelines rather than relying solely on data-driven learning. This addresses a real limitation of current deep learning models, particularly their inability to adapt to evolving clinical protocols without retraining. 2.The formulation (e.g., CTV = dilation of GTV minus OARs) closely reflects actual clinical reasoning, making the method intuitive and potentially more trustworthy. 3.The method achieves performance comparable to a strong supervised baseline (nnU-Net with GTV prior), despite not requiring task-specific training for CTV delineation. 4.The clinical assessment is a significant strength and provides valuable insight beyond standard segmentation metrics. 5.The explicit, human-readable planning steps improve transparency and enable rapid adaptation to new guidelines without retraining.

  • 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.While the framework does not require CTV-specific training, it relies on pre-trained OAR segmentation models and a large language model. This should be more clearly articulated to avoid overstating the training-free nature of the approach. 2.The method critically depends on pre-trained OAR segmentation models, yet the specific models used, their training data, and their performance are not described. Since OAR delineation directly influences the final CTV, this dependency should be explicitly characterized. 3.The evaluation is performed on a small test set (n=8), and the clinical assessment involves only two physicians. This limits the robustness of both quantitative and subjective conclusions. 4.The Likert-based evaluation aggregates ratings across a small number of cases and ratersr. Additionally, the absence of ground-truth contours in the evaluation makes it difficult to contextualize physician preferences. 5.While the method is applied to alternative guidelines and anatomical sites, quantitative evaluation beyond the primary esophageal task is limited. For example, prostate results are reported only via tool-call metrics rather than volumetric accuracy. 6.As acknowledged by the authors, the framework may be susceptible to misinterpretation or hallucination by the LLM, which could lead to incorrect delineation steps.

  • 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 does not provide sufficient information for 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

    The paper presents an innovative and clinically relevant direction for guideline-aware radiotherapy planning. To further strengthen the work, the authors could improve clarity around model dependencies, expand clinical validation, and provide more detailed evaluation of generalization across anatomical sites. Including ground-truth contours in the physician evaluation would also help contextualize subjective ratings.

  • 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 introduces a novel and clinically meaningful paradigm for target volume delineation that directly incorporates clinical guidelines into the contouring process. The approach demonstrates competitive performance with supervised methods while offering improved interpretability and adaptability. Despite limitations in dataset size, evaluation detail, and clarity of certain claims, the overall contribution is significant and has the potential to influence future research directions in radiotherapy planning.

  • 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.

    The rebuttal satisfactorily addresses several of my concerns. In particular, the authors clarify that zero-shot refers to the absence of CTV-specific training rather than complete independence from pretrained models, provide additional details on the OAR segmentation model, and appropriately acknowledge the limited prostate generalization claim. The concerns regarding small validation scale and limited physician evaluation remain, but these are reasonable limitations given the novelty and clinical relevance of the proposed paradigm. I therefore maintain my positive recommendation.



Author Feedback

We thank the reviewers for their thoughtful and constructive feedback. We are grateful for the recognition of OncoAgent’s novelty as the first guideline-aware AI agent for CTV delineation and its clinical relevance. We address the main concerns below.

1.Implementation Details and Reproducibility (R1, R3) We will add the following sentence after the implementation paragraph in Sec. 3.1: “GPT-5.2 was configured with temperature=1.0 and reasoning_effort=none; a modified 3D U-Net trained on a total of 2,870 CT and MR volumes with Dice 0.80–0.98 was used as an OAR segmentation model.” The full system prompt, JSON schemas, and execution code will be released on GitHub upon acceptance.

2.Analysis on Hallucination Frequency (R1, R3) OncoAgent’s hallucination risk falls into two categories: invalid tool-call plans (structural) and misinterpreted guidelines (semantic). For structural hallucination, the reported 1.13 calls/case in Sec. 3.1 implies that ~87% of plans pass schema validation on first generation while ~13% trigger self-refinement; all executed plans are structurally valid by construction. For semantic correctness, Tool Call F1 of 0.73–1.00 across heterogeneous esophageal guidelines (Table 3) provides indirect evidence. Regarding output stability, we empirically observed that the schema-validation and self-refinement loop drove repeated queries on the same case toward consistent, schema-valid plans. We will clarify these points in Sec. 3.1. 3.Task Formulation (R2, R3) By “zero-shot” we mean no CTV-specific training is needed; the framework still relies on a pre-trained LLM and OAR segmentation models. To avoid overstatement, we will rephrase terms such as “training-free” as “without any CTV annotation”.

4.Value of LLM Compared to Rule Engine and Complexity of CTV Decision Rules (R2) The contribution of the LLM is not measured by the number of LLM API calls, but by the elimination of per-guideline engineering cost. Target volume plans differ across guidelines—e.g., IJROBP follows a GTV→CTV→PTV pipeline with anatomy-aware CTV expansions, whereas CROSS expands GTV directly to PTV with fixed geometric margins. The value of LLM lies in autonomously adapting to different guidelines without any system-prompt or code modification, whereas rule engines would require hand-crafted parsers per guideline. Regarding CTV decision complexity, Equation 1 represents the computational skeleton shared across all guidelines, while guideline-specific complexity and patient-specific context are handled through LLM reasoning. For example, the LLM can dynamically determine which lymph node regions to include based on GTV location—cervical nodes for upper esophageal lesions, celiac nodes for lower. Furthermore, for cases requiring careful adjustment such as re-irradiation, our human-readable plan supports human-in-the-loop audit before contour generation.

5.Cross-Generalization (R2, R3) We acknowledge that the prostate Tool Call F1 of 0.64 indicates limited generalization to this site, and we will tone down the corresponding cross-site claims in Sec. 3.4.While this work focuses on esophageal cancer, the framework is designed to be extensible: incorporating site-specific features (e.g., OAR contours, target-construction methods) would enable broader applicability. We also clarify that Tool Call F1 is not a loose proxy: since the ground-truth calls were constructed by a human and execution is deterministic, matching sequence implies matching CTV contours.

6.Validation Scale (R2, R3) The current validation scale is acknowledged in our existing limitations, with larger cross-institutional, cross-guideline validation as a future direction. We note that, although ground truth per physician was unavailable, the expert contours in Fig. 3 served as a common visual reference during evaluation. Within this scope, OncoAgent demonstrates a shift from learning-from-data to reasoning-from-guidelines, evidencing a scalable and auditable pathway.



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.

    The paper receives 2 weak reject and 1 weak accept. All reviewers raise several valid major concerns. Rebuttal is invited.

  • 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.

    This paper still receives mixed reviews after rebuttal. One reviewer increases the score from negative to positive. Although this direction is interesting, I agree with R2 that, under the agent system, it is still not clear how GTV location is extracted, how upper/mid/lower esophageal involvement is determined, how margin ranges and physician preferences are operationalized, and whether such decisions are correct at the case level. The current quantitative performance is inferior to GTV prior based nnUNet. Moreover, the comparing method is not state-of-the-art, as there is more sophisticated clinical target volume segmentation utilizing GTV, OARs in esophageal and head & neck cancers. The evaluation dataset size is also small. Hence, I lean to rejection of this work.



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.

    The LLM driven agent framework for delineation while considering the organs at risk is novel. Authors carefully addressed most of reviewers’ concerns in the rebuttal. However, some of the limitations including lack of cross-site validation, limited data size, and how model is refined for individual patient-specific decision making must be clarified in the final paper.



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.

    This paper presents an LLM-based agentic framework that translates free-text clinical guidelines into an executable target volume contouring plan. I agree with the reviewers’ consensus regarding the clinical relevance of this work. After the rebuttal phase, the authors generally addressed the primary technical concerns raised during the review process. While some concerns remain (such as the issues regarding the small dataset scale, system reliability, and reproducibility), I think the novelty and clinical relevance of the proposed paradigm still outweigh the limitations, and it could provide valuable insights and inspiration for future research in automated radiotherapy planning. Therefore, I recommend accept.



back to top