The use of large language models in detecting Chinese ultrasound report errors

Abstract

This retrospective study evaluated the efficacy of large language models (LLMs) in improving the accuracy of Chinese ultrasound reports. Data from three hospitals (January-April 2024) including 400 reports with 243 errors across six categories were analyzed. Three GPT versions and Claude 3.5 Sonnet were tested in zero-shot settings, with the top two models further assessed in few-shot scenarios. Six radiologists of varying experience levels performed error detection on a randomly selected test set. In zero-shot setting, Claude 3.5 Sonnet and GPT-4o achieved the highest error detection rates (52.3% and 41.2%, respectively). In few-shot, Claude 3.5 Sonnet outperformed senior and resident radiologists, while GPT-4o excelled in spelling error detection. LLMs processed reports faster than the quickest radiologist (Claude 3.5 Sonnet: 13.2 s, GPT-4o: 15.0 s, radiologist: 42.0 s per report). This study demonstrates the potential of LLMs to enhance ultrasound report accuracy, outperforming human experts in certain aspects.