A fact-ledger framework for human-in-the-loop radiology report restyling.
放射科報告風格改寫的事實層框架:先抽取可審核的臨床事實,再由醫師修改事實層,最後交給 LLM 依指定風格產生新報告。
RadRestyle is a workflow framework for transforming a prior radiology report into an updated report draft without treating the language model as the source of truth. The central object is a fact ledger: a structured, editable list of clinical facts, evidence snippets, actions, and safety gates.
The framework is intentionally model-agnostic. It can be used with Codex, Claude Code, local agents, or another LLM environment. This public repository does not include private report-style rules, clinical data, physician identifiers, institution-specific templates, or UI source code from any deployment.
Radiology report rewriting is not just a text style-transfer task. A clinically useful workflow needs to separate:
- what was stated in the previous report,
- what has changed in the current encounter,
- what should be kept, revised, added, or removed,
- how the final report should sound,
- and what must be reviewed by a clinician before use.
RadRestyle formalizes that separation.
flowchart LR
A["Prior report text"] --> B["Fact extraction"]
B --> C["Editable fact ledger"]
C --> D["Current-encounter fact updates"]
D --> E["Safety validation"]
E --> F["Style rendering"]
F --> G["Clinician-reviewed draft"]
docs/architecture.mddescribes the end-to-end framework.docs/fact_ledger_schema.mdexplains the fact ledger object model.schemas/fact_ledger.schema.jsonprovides a JSON Schema starter.skills/generic/contains generic agent skill templates.prompts/contains model-agnostic prompt templates.examples/synthetic/contains fully synthetic examples for testing the workflow.docs/safety_validation.mddescribes hallucination controls and review gates.docs/ui_principles.mdrecords interaction principles without shipping any private UI implementation.
This repository deliberately excludes:
- real patient reports,
- medical record numbers or accession numbers,
- physician identifiers,
- institution-specific reporting templates,
- private report-style skills,
- current production UI source code,
- local small-model weights or inference code.
- Paste a prior report into an agent environment.
- Extract a structured fact ledger.
- Let a clinician review and modify the fact ledger.
- Add current-encounter findings.
- Render a new report using a private style skill or local reporting convention.
- Run safety checks before clinical use.
RadRestyle is a research and workflow framework. It is not a medical device and does not replace clinician review. Generated reports must be verified against the source imaging, clinical context, and institutional policy before use.
RadRestyle 的核心想法是不要讓 LLM 直接「憑空改寫」報告,而是先建立一個可檢查、可修改的事實層。醫師可以在事實層決定哪些資訊要保留、修正、加入或移除,之後再讓 LLM 依照私有的報告風格規則產生新報告。公開版本只提供框架、schema、通用 skill/prompt template 和合成範例,不包含任何真實病人資料或特定醫師的風格規則。
MIT License. See LICENSE.