1 h 3 min
Seminar, Hosted by Stanford MedAI
Collaboration between clinicians and vision–language models in radiology report generation
Google DeepMind
Recording
Abstract
Radiology informs the diagnosis, treatment and management of patients through medical imaging and radiology reports, but a global shortage of radiologists restricts access to expert care and creates heavy workloads that lead to delays and errors in clinical decisions. This talk shows how a specialist language model for radiology, fine-tuned from a vision-language model, can generate chest X-ray reports that a panel of radiologists rated as equivalent or preferred to radiologist-written reports in normal outpatient settings. For reports that contain errors, it presents an assistive setting in which clinicians and AI collaborate to write radiology reports, indicating new possibilities for clinical utility.