BCER: An Agent Execution Framework for Long-Running MRI Analysis Workflows

BCER (Bounded Cerebellum Execution Runtime) is an agent framework that separates planning and execution. It achieves reliable execution and fault recovery for long-running MRI analysis workflows through a bounded reflection mechanism, addressing issues like tool invocation failures and parameter errors in traditional single planning-execution loops. It supports multi-domain MRI analysis tools and has been open-sourced as supplementary code for a MICCAI paper.

MRI data processing involves multi-step complex workflows with extremely high accuracy requirements. Traditional automation methods' single planning-execution loops face issues like tool invocation failures and parameter transfer errors, some of which require manual intervention. BCER aims to achieve reliable execution by separating planning and execution and using a bounded reflection mechanism.

BCER adopts a four-layer architecture:

• Brain Layer: Constrained planning, generating constraint sketches based on predefined tools;
• Compiler Layer: Converting to a validated DAG;
• Cerebellum Layer: Deterministic execution, binding parameters and tool invocations;
• Extremity Layer: Encapsulating 21 domain-specific tools (covering prostate, brain, and cardiac workflows). The Reflector has a two-tier recovery system: Tier1 (rule layer handling known faults) and Tier2 (LLM layer handling complex issues, clearly distinguishing recoverable/unrecoverable faults).

The tool registry covers three major domains:

• Prostate workflow (fastMRI dataset);
• Brain workflow (BraTS 2021 dataset);
• Cardiac workflow (ACDC/CMRxRecon datasets). Tools are divided into three layers by complexity: basic tools (pure Python), inference tools (PyTorch/MONAI), and reconstruction tools (specialized k-space libraries), supporting selective installation.

Eight fault types are defined:

• Recoverable deterministic faults (Token variation, path variation);
• Recoverable semantic faults (parameter omission, semantic swap, spatial mismatch);
• Unrecoverable faults (modality missing, range violation, timeout). Evaluation supports four comparison strategies: BCER-only, ReAct baseline, ReAct + symbolic binding, and full ReAct + binding + reflector, quantifying the robustness contribution of each component.

Includes:

• Flexible scheduling modes (in-process/auto/subprocess);
• Environment isolation (complex tools in independent conda environments);
• Reproducibility design (YAML configuration, results bound to versions).

Academic contributions:

• Methodology: A new paradigm combining planning and execution;
• Empirical research: Fault injection experiments provide benchmark data;
• Open-source ecosystem: Lowering research barriers. Application scenarios: Used as a workflow orchestration engine in hospital imaging departments, AI companies, and research institutions.

Open-sourced under the MIT license. Quick start steps:

1. Clone the repository: git clone https://github.com/Albertlongzi/BCER.git
2. Create environment: conda env create -f envs/base.yml
3. Activate environment: conda activate bcer-base
4. Install: pip install -e .
5. Smoke test: python -m benchmark.smoke Verify installation correctness; complete deployment guidance is provided in the documentation.