RadEx: A Framework for Structured Information Extraction from Radiology Reports Based on Large Language Models.
Authors
Affiliations (3)
Affiliations (3)
- Bern University of Applied Sciences, Institute for Patient-centered Digital Health, Biel/Bienne, Switzerland.
- University of Geneva, PhD School of Life Sciences, Geneva, Switzerland.
- ID SUISSE AG, St. Gallen, Switzerland.
Abstract
Radiology reports are typically written as unstructured free text, limiting their potential for applications like automatic analysis, clinical trial matching, and outcome prediction. The adoption of structured reporting remains low due to established workflows, resource constraints, and concerns about information loss. This study presents RadEx, an end-to-end framework for structured information extraction from radiology reports. RadEx consists of 15 software components and ten artifacts, covering the full process from data annotation to information extraction. The framework enables clinicians to define domain-specific information and create report templates. It supports both generative and encoder-based models, while decoupling template filling from model inference to allow flexible model updates. By standardizing artifacts and modularizing components, RadEx facilitates implementation, maintenance, and interoperability.