Our team investigated the feasibility of using GPT-4, a cutting-edge natural language processing model, to automate the conversion of free-text radiology reports into structured templates. We evaluated GPT-4’s performance on 170 detailed CT and MRI reports in English and 583 chest X-ray reports in German. The model successfully transformed all free-text reports into valid JSON files without losing accuracy or indicating additional findings. Moreover, GPT-4 outperformed the existing state-of-the-art in detecting pathological findings and therapeutic devices in the German chest X-ray dataset.
Our findings, published in Radiology (https://doi.org/10.1148/radiol.230725), highlight the potential of generative models like GPT-4 to structure vast amounts of unstructured data in medical databases with minimal programming effort. While there might be limitations related to data privacy when using third-party models, our proof-of-concept study demonstrates the transformative potential of GPT-4 in radiology and healthcare. We believe this technology can facilitate research, data sharing, and empower non-computer scientists to lead database structuring efforts within our field.
AIAH Lab 