Comparing GPT and Claude visual language models in radiology

Abstract

Vision Language Models (VLMs) are emerging tools in radiology, with the potential to aid clinical workflows by identifying anatomical regions and imaging modalities from complex datasets. Recent advancements in models like GPT and Claude show promise for improving diagnostic efficiency, particularly in recognising anatomical structures, detecting fractures, and classifying imaging modalities. However, previous studies highlight limitations in the reliability and diagnostic accuracy of these models, especially in nuanced pathology identification. This study seeks to address these gaps by assessing the proficiency of popular publicly available GPT and Claude models in key diagnostic tasks. Improved model performance has the potential to enhance radiology workflows, optimise resource utilisation, and support clinical decision-making, but further evaluation is required to establish their clinical impact.