List of Papers Browse by Subject Areas Author List
Abstract
Recent medical VLM and agent studies are often evaluated on 2D images or short tool-calling tasks, while practical MRI analysis requires long, interdependent workflows over 3D/4D volumetric data. In such workflows, reactive tool-calling agents are vulnerable to cascading failures caused by incorrect intermediate references, incompatible tool arguments, and weak control over cross-step dependencies. We present BCER (Brain–Cerebellum–Extremity–Reflector), a controller architecture for reliable long-horizon MRI workflow execution. BCER separates high-level planning from execution and supports bounded local recovery. We evaluate BCER on a multi-organ MRI benchmark spanning brain, prostate, and cardiac tasks with both short and long-chain workflows, under matched task contracts across controller variants and multiple backbone models. Compared with reactive baselines, BCER consistently improves end-to-end execution, with the largest gains on long-chain workflows. BCER also supports auditability by preserving explicit links between final outputs and intermediate artifacts and measurements. Code and benchmark are available at https://github.com/Albertlongzi/BCER.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0902_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/Albertlongzi/BCER
Link to the Dataset(s)
N/A
BibTex
@InProceedings{LonZiy_BCER_MICCAI2026,
author = { Long, Ziyang AND Li, Xinqi AND Chen, Junzhou AND Gao, Yifan AND Li, Debiao AND Yang, Hsin-Jung},
title = { { BCER Agent: Reliable Long-Horizon MRI Workflow Execution via Compilation, Artifact Binding, and Bounded Local Recovery } },
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 paper proposes BCER, a hierarchical agent controller that separates planning from execution to handle long, interdependent MRI analysis workflows over 3D/4D data. It claims improvements over reactive baselines on benchmark datasets.
- 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 identification of three specific failure modes — incorrect intermediate references, incompatible tool arguments, and weak cross-step dependency control — makes the contribution significant.
Repair-and-retry instead of full workflow restarts” is a concrete and distinguishing architectural commitment.
Preserving traceable links between final outputs and intermediate artifacts has direct clinical relevance — regulatory and research pipelines require provenance.
Matched task contracts across controller variants and multiple backbone models suggests an attempt at systematic ablation.
- 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 core ideas are clinically motivated and technically sound in direction — but the paper has significant gaps in architectural specificity, metric rigor, and terminological precision.
Plan-execution separation with local recovery has precedent in classical HTN planning, DEPS, and reflexion-style architectures. It is not very clear how BCER is different from them or provides a comparative analysis.
Success Rate (SR) is described as “satisfying the full task-specific contract,” but the contract is never formally defined. Is SR binary per workflow or averaged across steps? For 3D/4D outputs like segmentation masks or registration fields, what constitutes success—Dice above a threshold, anatomical plausibility, or exact match?
Similarly, Task Completion Rate (TCR) is “the number of workflow completion steps with required contract outputs”. Does that mean steps that produced any output, or steps that produced correct outputs? With reference to total planned steps, total executed steps, or total reference workflow steps? Could a high TCR produce clinically meaningless results?
- Please rate the clarity and organization of this paper
Poor
- 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?
Lack of clarity in evaluation metrics is a major factor
- 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
The main contribution of the paper is a reliable execution control for long-horizon MRI pipelines. It proposes an integrated system where the brain orchestrates the analysis plan, cerebellum creates executable workflow graph with symbolic artifact binding, extremity are the standardized MRI tool library and Reflector, an error corrector that retries only failed nodes with minimal patches instead of restarting entire workflows.
- 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 authors provide clear problem statement and describe multiple tasks across multiple datasets. Via ablation studies, their evaluation shows that the performance improvement comes from the execution control.
- 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 components of the system are not novel by themselves, though the integration and evaluations are meaningful making the work a bit incremental. Comparisons with expert evaluations are not provided.
- 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 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.
(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 reads like a strong engineering project rather than a novel analytic algorithm. Furthermore, no direct comparison with expert generated outcomes were provided.
- 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
Review #3
- Please describe the contribution of the paper
This paper introduces BCER (Brain–Cerebellum–Extremity–Reflector), a structured controller architecture for reliable long-horizon MRI workflow execution. Unlike conventional reactive LLM-based agents, the framework separates high-level planning, execution, tool interaction, and failure recovery. The system treats MRI analysis as a multi-step workflow problem rather than isolated tasks and introduces symbolic artifact binding and bounded local recovery to improve execution reliability. The paper also proposes a task-contract benchmark across multiple MRI domains and evaluates performance using success rate and task completion rate. The results show strong improvements in long-chain workflows compared to reactive 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.
The paper addresses a highly relevant and underexplored problem in medical AI, namely reliable execution of multi-step MRI workflows. The architecture is well-designed with clear modular separation between planning, execution, and recovery. The integration of symbolic artifact binding is particularly useful for handling intermediate dependencies in medical pipelines. The experimental results are strong and clearly demonstrate large improvements over baselines, especially in long-chain tasks where reactive agents fail. The benchmark is well-constructed with multiple datasets and task types, and the use of task contracts is a meaningful evaluation strategy. The work is also practically relevant, with emphasis on auditability and traceability which are important for clinical adoption.
- 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 novelty is primarily at the system integration level rather than introducing fundamentally new methodology. Many components such as ReAct-style agents, reflection mechanisms, and modular tool frameworks already exist in prior work. The paper lacks mathematical formulation or theoretical guarantees regarding stability or convergence. Comparisons are limited mostly to ReAct variants, and stronger or more recent baselines in medical agent systems are missing. The framework assumes a curated and standardized tool library, which may not reflect real-world clinical environments. The dataset size is relatively limited and evaluation is constrained to predefined task contracts. There is no clinical validation or human-in-the-loop evaluation to demonstrate real-world impact.
- 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.
(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 addresses an important and practical problem in medical AI, namely reliable execution of long-horizon workflows. The proposed BCER framework is well-designed and demonstrates strong empirical improvements, especially for long-chain tasks where existing approaches fail. However, the novelty is largely at the system integration level, and the lack of theoretical contribution and limited baseline comparisons reduce its overall strength. Despite these limitations, the practical value and clear improvements make it slightly above the acceptance threshold.
- 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
Thank the reviewers and Area Chair for the constructive comments and for recognizing the importance of reliable long-horizon MRI workflow execution. Here, we respond to the three main clarification points raised by the reviews and AC.
First, regarding SR and TCR, we agree that the submitted definition was too concise. Each case is evaluated against a fixed task contract specifying the input, goal, allowed tools, required workflow milestones, and required deliverables. Success Rate (SR) is binary at the case/workflow level: a run is successful only if all required milestones and deliverables are completed and pass the corresponding contract validators. Task Completion Rate (TCR) is a milestone-level measure: it is the fraction of required contract milestones whose specified outputs are produced and validated, using the reference contract milestones as the denominator rather than attempted or generated steps. Thus, arbitrary files or free-form answers do not count toward TCR. TCR measures partial contract completion, whereas SR measures full end-to-end contract satisfaction. These metrics evaluate controller-level execution reliability rather than clinical diagnostic accuracy or the intrinsic performance of individual MRI algorithms.
Second, regarding related architectures, we do not claim that plan/execution separation, tool use, or reflection is independently new. BCER instead integrates these ideas for long-horizon volumetric MRI workflows, where intermediate images, masks, measurements, spatial metadata, and cross-step dependencies must be handled as first-class execution objects. Unlike general HTN or DEPS-style planning, BCER does not search over an open-ended symbolic domain or treat an LLM plan as directly executable. The Brain produces a constrained plan sketch, while the Cerebellum compiles it into a typed, case-scoped workflow graph with dependency resolution and symbolic artifact binding. Unlike broad Reflexion/Self-Refine-style revision, the Reflector performs bounded local repair only on failed nodes or small affected sub-workflows, preserving valid upstream results. Compared with medical agents focused mainly on diagnosis reasoning, VQA, or short tool use, BCER targets reliable execution and provenance of multi-step MRI processing pipelines.
Third, regarding clinical validation and expert-generated outcomes, we agree that this is an important limitation. This study evaluates controller reliability under fixed MRI workflow contracts; it is not a prospective clinical validation study and does not claim to replace expert-generated reports or clinician judgment. Expert knowledge enters through task definitions, public benchmark annotations when available, predefined deliverables, and rule/tool interfaces. The intended benefit is complementary to expert review: BCER links final outputs to intermediate artifacts, measurements, tool calls, error traces, and recovery decisions, supporting reproducible and auditable clinical research workflows. Prospective clinician-in-the-loop validation remains necessary before clinical deployment.
As noted above, we will incorporate these clarifications into the camera-ready version.
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 presents a novel AI agent for planning, executing, and verifying long-horizon MRI workflows. As acknowledged by all reviewers, this paper tackles an important and practical problem in MRI analysis. The motivation is well-stated, and the framework is evaluated on multiple tasks (short- and long-tasks) across multiple MRI datasets spanning three different body regions. Overall, the relevance of this work to the MICCAI community encourages my decision to provisionally accept the paper. However, I strongly advise the authors to address the few weaknesses raised by the reviewers:
- Clarify the evaluation metrics. R1 found unclear descriptions of metrics like “Success Rate” and “Task Completion Rate”.
- Position the contribution with respect to similar methods pointed out by reviewers: classical HTN planning, DEPS, reflection-style architectures, and other AI medical agents.
- Absence of clinical validation or human-in-the-loop evaluation. I suggest that the authors add a few sentences to better highlight the framework’s differences and benefits relative to expert-generated outcomes.