[论文解读] EndoExtract: Co-Designing Structured Text Extraction from Endometriosis Ultrasound Reports

Haiyi Li, Yiyang Zhao+6 more|arXiv (Cornell University)|Jan 26, 2026

Radiology practices and education被引用 0

一句话总结

EndoExtract 是一个与临床医生共同设计的本地部署的基于大语言模型的系统，用于从子宫内膜异位症超声报告中提取结构化数据，优先实现解释性字段复核、自动证据高亮和批量化验证。

ABSTRACT

Endometriosis ultrasound reports are often unstructured free-text documents that require manual abstraction for downstream tasks such as analytics, machine learning model training, and clinical auditing. We present extbf{EndoExtract}, an on-premise LLM-powered system that extracts structured data from these reports and surfaces interpretive fields for human review. Through contextual inquiry with research assistants, we identified key workflow pain points: asymmetric trust between numerical and interpretive fields, repetitive manual highlighting, fatigue from sustained comparison, and terminology inconsistency across radiologists. These findings informed an interface that surfaces only interpretive fields for mandatory review, automatically highlights source evidence within PDFs, and separates batch extraction from human-paced verification. A formative workshop revealed that extbf{EndoExtract} supports a shift from field-by-field data entry to supervisory validation, though participants noted risks of over-skimming and challenges in managing missing data.

研究动机与目标

• Characterize workflow pain points in endometriosis report abstraction through contextual inquiry.
• Develop an on-premise LLM-based extraction system tailored to clinical data review needs.
• Design an interface that shifts data entry from field-by-field capture to supervisory validation.
• Enable automatic evidence highlighting and semantic normalisation to support verification.
• Evaluate the approach through a formative workshop with domain experts to assess adoption and workflow impact.

提出的方法

• Contextual inquiry with research assistants to identify trust asymmetries and verification practices.
• Development of EndoExtract with a gpt-oss-20b backend deployed on-premises via Ollama for privacy.
• Selective review surface exposing only five interpretive fields for mandatory human review while automating 150+ numeric fields.
• Automatic highlighting of source evidence within PDFs to support verification.
• Batch processing of up to 5,000 PDFs with paced review and change tracking.
• Semantic normalisation of terminology to reduce cognitive load and improve consistency.

实验结果

研究问题

• RQ1 How do workflow factors and trust asymmetries affect data extraction from endometriosis ultrasound reports?
• RQ2 Can an on-premise LLM-based system with selective review surfaces improve efficiency and maintain data quality in clinical text abstraction?
• RQ3 Does evidence highlighting and batch-paced verification support more reliable supervisory validation of extracted data?
• RQ4 What design principles best support verification-centred human-AI collaboration in clinical information extraction?
• RQ5 What are the perceived risks and adoption barriers for an AI-assisted clinical data abstraction interface?

主要发现

• Trust asymmetry drives selective review for interpretive fields while automating numerical fields.
• Automatic evidence highlighting eliminates repetitive manual annotation and supports verification.
• Batch processing enables paced review and reduces fatigue without sacrificing traceability.
• Semantic normalisation reduces terminology variation and improves downstream consistency.
• Formative workshop suggests a shift from data entry to supervisory validation with attention to potential over-skimming and missing data.