Large Language Models for Structured Reporting in Radiology: Past, Present, and Future

Introduction/Context: Structured reporting (SR) has been a topic of discussion for decades as a way to standardize and improve the quality of radiology reports. Although there is evidence that SR can reduce errors and enhance guideline adherence, widespread adoption remains limited. Large language models (LLMs), transformer-based and trained on large volumes of text, have emerged as a promising solution to automate and facilitate structured radiology reporting. This narrative review provides an overview of LLM-based SR, while also discussing limitations, regulatory challenges, and future applications.

Key Points/Findings/Recommendations:

1. Application of LLMs in Structured Reporting: Studies focus on models such as GPT-3.5 and GPT-4, showing promising results in converting free-text radiology reports into structured formats, including multilingual capabilities.
2. Main Advantages:
   • Consistency and improved standardization of reports.
   • Potential to reduce errors and enhance comprehensiveness.
   • Streamlining the use of predefined templates and formats.
3. Multilingual and Translational Capabilities: Research shows that several LLMs can process reports in different languages, promoting broader global collaboration.
4. Technical Limitations: Hallucination (fabricated content), terminology inconsistencies, and misinterpretations remain hurdles to large-scale adoption.
5. Regulatory and Privacy Challenges: Proprietary model opacity and a lack of regulated pathways pose difficulties for safely integrating these systems into clinical practice.

Implications for Practice:

• Automation and Efficiency: LLMs can streamline the reporting process, reduce typing effort, and facilitate standardized descriptions, offering efficiency gains.
• System Integration: Incorporation of LLMs into radiology systems (PACS, RIS) could help with documentation and report formatting, improving communication across teams.
• Broader Clinical Perspective: Standardized reports may enhance information sharing and potentially patient safety, especially in multidisciplinary settings.

Limitations/Considerations:

• Hallucination: Even advanced models may generate inaccurate or fictitious content unrelated to the evidence.
• Lack of Transparency: Proprietary models often do not disclose their training data or algorithms, hindering external validation.
• Evolving Regulations: The European Union and other countries are developing AI-specific legal frameworks. Meeting safety, reliability, and transparency requirements is critical for clinical use.
• Availability of Open-Source Models: While open-source models offer more regulatory flexibility, they still require significant validation and refinement for clinical application.

Conclusion: LLMs have significant potential to transform structured radiology reporting, offering increased efficiency and accuracy. However, regulatory issues, model opacity, and current technical limitations must be addressed before these tools can be safely and effectively integrated into clinical practice. Future research should explore clinical acceptance of LLM-generated reports, compare them to radiologist-produced reports, and investigate how these models can be best integrated into existing systems.

Sources and Links:

• Busch, F., Hoffmann, L., dos Santos, D.P. et al. Large language models for structured reporting in radiology: past, present, and future. Eur Radiol (2024). https://doi.org/10.1007/s00330-024-11107-6