Accuracy of the Image Interpretation Capability of ChatGPT-4 Vision in Analysis of Hess Screen and Visual Field Abnormalities.

Abstract

OpenAI, the owner of ChatGPT, publicly released the GPT-4 Vision in September 2023. This multimedia chatbot has the capability to receive and analyze various images presented to it by the user. We assessed the accuracy of its interpretation of 2 of the images commonly used in neuro-ophthalmology, namely Hess screen and automated visual field images. We separately uploaded typical images of 5 abnormal Hess screen charts related to third, fourth, and sixth cranial nerve palsy, Brown syndrome, and inferior orbital wall fracture with entrapment of the inferior rectus muscle. Likewise, 5 classic images of automated visual field grayscale maps related to lesions of the optic nerve, the chiasma, the optic tract, the optic radiations, and the occipital lobe were presented. The chatbot was instructed to select the best option among the 5 choices presented in each question. The GPT-4 Vision was able to select the right choice in 2/5 questions on Hess screens and 3/5 of the visual field questions. Despite selection of the correct option, qualitative evaluation of GPT-4 responses revealed flawed analysis of certain aspects of some image findings, such as the side of involvement or the misinterpretation of the physiologic blind spot as a central scotoma. The performance of GPT-4 Vision in the interpretation of abnormalities of Hess screen and visual field involvement was highly variable, even with simple typical cases of classic disorders. As the chatbot's image recognition is currently evolving, its capacity to accurately interpret ophthalmologic images is still limited at this time.