List of Papers Browse by Subject Areas Author List
Abstract
Medical landmark detection varies significantly across modalities and clinical protocols, posing unique challenges for developing universally applicable models. Conventional approaches typically entail training bespoke architectures per dataset or relying on general-purpose heuristics with fixed rules. However, these methods suffer from limited scalability and lack the semantic reasoning required to adapt to complex anatomical variations. To address these limitations, we propose AutoLand, a data-efficient autonomous workflow for landmark detection in medical images, leveraging a Vision Language Model (VLM) as a medical analysis agent. Specifically, AutoLand integrates a Physics Fingerprint Extraction module to extract precise physical characteristics, ensuring the agent aligns with modality-specific data properties. Subsequently, a Medical Deep Learning Library is queried during reasoning to retrieve specialized network architectures and optimization protocols, enabling the formulation of expert-level training strategies with only a few samples. This workflow enhances robustness and adaptability while reducing manual intervention. Experiments on three datasets show that AutoLand significantly outperforms state-of-the-art methods. Moreover, ablation studies confirm that our agentic paradigm effectively overcomes the limitations of fixed baselines through semantic reasoning. The source code is available at https://github.com/ChiZai123456/AutoLand.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0716_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{LiZiy_AutoLand_MICCAI2026,
author = { Li, Ziyi AND Li, Jiarui AND Shi, Yulong AND Jiang, Xue AND Qi, Shouliang AND Liu, Ting AND Qi, Lin AND Qian, Wei},
title = { { AutoLand: A Data-Efficient Automated Agentic Workflow for Universal Medical Landmark Detection } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16878},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The main contribution is an agent-guided workflow that formulates dataset-specific training pipelines for medical landmark detection from limited annotations, by combining modality-aware physics fingerprints with a constrained medical deep learning library.
- 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 proposes an interesting and well-motivated direction. (2) Automating per-dataset configuration for landmark detection could reduce manual tuning effort and expert dependence, which is relevant for real medical settings where datasets are limited and heterogeneous. (3) The physics fingerprint is a simple but practical idea. It provides the agent with compact modality- and acquisition-related information that may help guide more dataset-specific strategy selection. (4) The cross-dataset evaluation is reasonably broad. Validation on cephalometric X-ray, chest X-ray, and cardiac MRI helps support the claim that the framework is not tied to a single anatomy or modality. (5) The paper is clearly motivated and generally 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.
(1) The agentic framework is not described in enough detail, especially the prompt design, constraint mechanism, and strategy parsing. (2) The bridge from generic architecture selection to landmark detection is unclear, including the prediction head and task-specific design details. (3) Baseline fairness is unclear, as tuning budgets, hyperparameters, augmentations, and model settings are not standardized or sufficiently described. (4) The comparison does not isolate the value of the agent itself, since no non-agent AutoML baseline is included. (5) The data-efficiency claim is not fully validated, because performance is only reported under a fixed 1:1:8 split. (6) Potential patient-level leakage is not clearly ruled out, particularly for the cardiac MRI dataset. (7) Practical feasibility remains unclear, including compute cost, latency, and the reliability of the agent-generated strategies.
- 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?
The idea is relevant, timely, and potentially impactful, and the ablation studies suggest that both the physics fingerprint and the curated medical DL library contribute meaningfully to performance. However, my overall score is limited by several weaknesses in the current manuscript, including unclear dataset and experimental pipeline details, baseline comparisons that are not yet fully convincing, and reproducibility gaps in the description of the framework. Clarifying these aspects would substantially improve the quality and credibility of the work.
- 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
This paper proposes AutoLand, a data-efficient automated agentic workflow for universal medical landmark detection. It takes a large vision-language model (VLM) as a medical analysis agent to automatically synthesize dataset-specific training strategies, integrating a Physics Fingerprint module to extract modality-acquisition related data properties and a curated Medical Deep Learning Library for executable pipeline construction. Evaluations on cephalometric X-rays, chest X-rays, and cardiac MRI show superior performance over mainstream baselines, and ablation studies verify the effectiveness of the two core components. The work provides a new agentic paradigm for reducing manual tuning in universal medical landmark detection.
- 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.Technical novelty: Using VLM as a medical analysis agent to generate executable training strategies from a constrained DL library is an innovative system-level attempt for automated universal medical landmark detection. The Physics Fingerprint effectively provides modality-aware cues to alleviate agent hallucinations. 2.Experimental validation: Evaluated on three typical medical imaging modalities with comprehensive baselines and solid ablation studies; both quantitative metrics and qualitative visualizations are provided to support the effectiveness. 3.Practical value: The workflow targets the pain point of manual dataset-specific tuning in few-sample scenarios, which has high application potential for clinical deployment of landmark detection models. 4.Presentation clarity: The pipeline is clearly decomposed into three modules with formalized equations, making the core design 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.
1.The Physics Fingerprint only covers basic statistics and can be extended with more clinical metadata 2.The reliance on Qwen3-VL-235B lacks computational overhead quantification. 3.The landmark detection formulation and agent guardrail mechanism need more detailed description. 4.Multi-seed experiments and statistical significance tests are suggested to enhance rigor. 5.Insufficient discussion and comparison with representative universal landmark detection works, such as (UOD: universal one-shot landmark detection).
- 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
This paper puts forward a novel agentic automation paradigm for universal medical landmark detection with solid experimental results and high clinical application potential. To further improve the paper, it is recommended to: 1) Add multi-seed experiments and statistical tests; 2) Supplement detailed engineering details of the agent and DL library; 3) Cite and compare with the following representative works on (universal) (one-shot) landmark detection: Uod: Universal one-shot detection of anatomical landmarks Pele scores: pelvic x-ray landmark detection with pelvis extraction and enhancement Learning to Localize Cross-Anatomy Landmarks in X-Ray Images with a Universal Model
- 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?
AutoLand proposes an innovative agentic automation workflow for universal medical landmark detection, which combines VLM reasoning, Physics Fingerprint and Medical DL Library to solve the problem of manual tuning in few-sample scenarios. The core idea is novel, the experimental results on three modalities are promising, and the work has clear clinical value. Although the paper has imperfections in experimental details, related work completeness and technical description, these issues can be well addressed through minor revisions. The core contribution is solid and meets the basic acceptance standard of MICCAI, so a weak accept is recommended.
- 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 main contribution of this paper is the proposal of an automated workflow for general medical landmark detection. It leverages a VLM as an agent to guide the selection of components and training strategies, effectively automating the workflow to handle different modalities and anatomical structures.
- 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 provides an interesting and original solution to general keypoint detection. Instead of relying on fixed architectures or general-purpose heuristics, it creatively uses a VLM agent to adaptively formulate strategies. It effectively addresses key challenges in the current field of general keypoint detection, such as sensitivity to data modality and anatomical variations. The method has been validated across different datasets, demonstrating its initial capability, and the overall execution of the paper is quite solid.
- 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.
General keypoint detection methods are highly sensitive to data modality and anatomical structures. While validated on several datasets, the evaluation lacks sufficient diversity. Datasets from different anatomical regions (e.g., publicly available hand X-ray, spine X-ray, and abdominal ultrasound datasets) should be included to fully prove the method’s universality. Lack of Transparency in Agent Strategy: It remains unclear how one can verify that the strategy selected by the VLM agent is indeed the optimal one. Additionally, the paper lacks discussion on failure cases, such as whether the method falls back to baseline models like U-Net under certain conditions. For the Cardiac dataset split, the authors state that 5,293 LAX and 15,024 SAX slices are obtained from 144 patients. Adjacent slices from the same patient share highly similar anatomical structures. There is no confirmation that the split was performed strictly at the patient level. Splitting these slices randomly across training, validation, and test sets will lead to severe data leakage.
- 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?
Overall, this paper presents a solid and interesting VLM-guided solution that addresses real challenges in general landmark detection. The major factors leading to my current score are the promising novelty of the workflow, offset by significant concerns regarding the evaluation. Specifically, the risk of data leakage in the Cardiac dataset is a critical issue that must be clarified. Furthermore, the lack of anatomical diversity in the evaluation datasets and the missing verification of the agent’s optimality prevent a higher score. If the data leakage concern is resolved and the evaluation is broadened, I would be willing to give a higher evaluation.
- 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
Author Feedback
Dear Area Chair and Reviewers,
We sincerely thank the Area Chair and all Reviewers for their constructive feedback. We are deeply encouraged by your recognition. We would like to briefly clarify the most critical concern and outline our revision plan: 1.Clarification on the Cardiac MRI Dataset Split: To address the concern regarding potential data leakage, we confirm that the 1:1:8 partition for the Cardiac MRI dataset was conducted strictly at the patient level. There is no patient-level data leakage. We sincerely apologize for the ambiguity in the paper and will explicitly state this patient-level split strategy in the camera-ready version. 2.Detailed Engineering Implementations: We appreciate the suggestions to include more details on prompt design, guardrail mechanisms, and computational overhead. However, due to strict page limits, we will comprehensively supplement all exact engineering details and source code in our open-source repository. 3.Generalization to Broader Datasets: We are currently refining our agentic workflow and expanding our experiments to include more diverse datasets, such as hand and spine datasets. Thank you once again for your invaluable guidance.
Best regards, Authors of Paper #716
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.
AutoLand introduces a novel agentic paradigm that uses a VLM together with a Physics Fingerprint and a curated medical DL library to automatically synthesize dataset-specific training strategies for universal medical landmark detection — a timely, well-motivated direction with clear clinical value. All three reviewers consistently affirm the originality of the workflow, the effectiveness of both core modules in ablations, and solid cross-modality validation on cephalometric X-ray, chest X-ray, and cardiac MRI; remaining concerns (cardiac patient-level split, broader baselines, agent reliability) are addressable in camera-ready.