Scholarlink

Big data research in neuroradiology relies on labeled datasets, often extracted manually from imaging in a time- and labor-intensive process. Although artificial intelligence (AI) advances automated analysis, security concerns limit the sharing of medical images. Radiology reports offer an alternative but require manual labeling. AI-assisted labeling could be beneficial, but privacy risks arise with cloud-based tools such as ChatGPT, and many AI models lack specialized development for neuroimaging. We propose SCOPE (Stroke COntent Parsing and Extraction), a novel approach that combines an open-source large language model (Llama-3.1) and retrieval-augmented generation (RAG) to extract stroke diagnosis labels from medical reports. This method generates labeled datasets linked to patient imaging data for clinical research. Leveraging a pre-trained Llama model with RAG eliminates the need for fine-tuning (i.e., re-training) and allows seamless data expansion. With an overall accuracy of 0.93, a sensitivity of 0.92, a specificity value of 0.96, and an F1-score of 0.95, SCOPE outperforms Llama-3.1 without RAG and Llama-3.1 with fine-tuning. The GitHub code is available: https://github.com/mumuaktar/SCOPE/ .

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

SCOPE: Label Extraction of Stroke Diagnosis from Unstructured Medical Reports Using Retrieval-Augmented Generation

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Mumu Aktar,
  • Gunjan Jindal,
  • Salome Lou Bosshart,
  • Alexander Stebner,
  • Pedro Paiva,
  • Mariana Bento,
  • Johanna Ospel,
  • Roberto Souza

摘要

Big data research in neuroradiology relies on labeled datasets, often extracted manually from imaging in a time- and labor-intensive process. Although artificial intelligence (AI) advances automated analysis, security concerns limit the sharing of medical images. Radiology reports offer an alternative but require manual labeling. AI-assisted labeling could be beneficial, but privacy risks arise with cloud-based tools such as ChatGPT, and many AI models lack specialized development for neuroimaging. We propose SCOPE (Stroke COntent Parsing and Extraction), a novel approach that combines an open-source large language model (Llama-3.1) and retrieval-augmented generation (RAG) to extract stroke diagnosis labels from medical reports. This method generates labeled datasets linked to patient imaging data for clinical research. Leveraging a pre-trained Llama model with RAG eliminates the need for fine-tuning (i.e., re-training) and allows seamless data expansion. With an overall accuracy of 0.93, a sensitivity of 0.92, a specificity value of 0.96, and an F1-score of 0.95, SCOPE outperforms Llama-3.1 without RAG and Llama-3.1 with fine-tuning. The GitHub code is available: https://github.com/mumuaktar/SCOPE/ .