Researchers from Germany evaluated whether multimodal large language models (LLMs) with audio capabilities could transform radiology reporting workflows, comparing GPT-4o and Claude Sonnet 3.5 against conventional speech recognition software. The study, published in European Radiology in August 2025, analyzed 480 reports from 80 multimodal imaging cases reported by two board-certified radiologists. LLM workflows demonstrated remarkable efficiency gains, reducing reporting time by more than half while requiring significantly fewer corrections and improving cost efficiency by 23%. Claude Sonnet 3.5 reports were rated highest in quality, while GPT-4o matched conventional report quality.
The findings confirmed that minimal audio input could enable complete radiological report generation, with both LLMs generating structured, high-quality reports more efficiently than conventional speech recognition methods. These results suggest that multimodal LLMs show promise for improving radiology reporting efficiency and quality through audio-based report generation, potentially transforming traditional dictation workflows while maintaining or improving report standards.