ChatGPT-Assisted Identification of Cerebrovasculature: A Comparative Analysis between an Anatomical Model and Digital Subtraction Angiography

Introduction: Angiography is a critical imaging modality in vascular neurosurgery, with its utility relying on a thorough understanding of neuroanatomy and cerebrovasculature. With the advancement of the ChatGPT-4 Artificial Intelligence (AI) model to include image processing, the question of its reliability in clinical medicine remains unconfirmed. We explore the accuracy of ChatGPT-4’s foundational neuroanatomy and its applications in vessel recognition in digital subtraction angiography (DSA).

Methods: DSAs collected from online sources, including 3 angiogram views, were given to ChatGPT-4 to identify image modality and specified vasculature. Major blood vessels that were identifiable on each view, as designated by a cyan arrow, were tested. Per angiogram view, each vessel was fed to ChatGPT-4 as 3 separate queries and prompted with the following, “what vessel is the cyan arrow pointing to and why is it that arrow?”

Results: ChatGPT-4 identified the correct imaging modality 100% of the time. Across a total of 39 queries, the model correctly identified the designated vessel on the angiogram 25.64% of the time. Notably, out of 39 queries, the model incorrectly identified the specified vessel as the anterior communicating artery, which was not visible on any image, 21% of the time, followed by the aorta (10%). Across a total of 6 queries on the circle of willis, the model showed 16.67% accuracy across 6 vessels. Further, the model identified each of the 6 different vessels as the anterior communicating artery. When asked to identify the general structure, it correctly identified the circle of willis.

Conclusion : Our study demonstrates that ChatGPT-4, while a significant advancement in the field of Artificial Intelligence, possesses immature functionality pertaining to anatomical identification of cerebrovasculature, irrespective of image type. Further research is needed to evaluate trends in the model’s recognition of cerebrovasculature in additional neuroimaging modalities, and compare the accuracy of results amongst these.