(VIRTUAL) Investigating Chat-GPT-4o's ability in diagnosing dermatopathological images

Track
Basic Science
Abstract

Large language models (LLMs) are continuously evolving, with OpenAI releasing its latest product version, ChatGPT-4o (GPT-4o), enabling quicker responses, detailed answers, and upgraded image and video analysis. Studies have shown that previous versions of GPT were unable to analyze dermatopathology images. Therefore, in this study, we investigated the upgraded image analysis capabilities of GPT-4o to determine if it can accurately diagnose dermatopathology images. To test its abilities, we provided the LLM with various dermatopathology images and asked it two standardized questions: 'What is the most likely diagnosis?' and 'What are the most common differential diagnoses?”. Forty hematoxylin and eosin (H&E) stained digital slides, deemed suitable by a trained dermatopathologist, were chosen either from a pathology dataset or from DermNet™, a free online dermatology database. Ten basal cell carcinomas (BCC), ten squamous cell carcinomas (SCC), ten compound melanocytic nevi (CMN), and ten invasive melanomas were included in this study. In our study, GPT-4o correctly diagnosed 20% of BCC images, 40% of SCC images, and 10% of CMN and melanoma images. In 70% of cases, GPT-4o included the correct diagnosis in the list of differentials. This study demonstrated that the new GPT-4o can analyze histology images. Although larger studies with more extensive datasets are needed to further evaluate its accuracy, rapid advancements in LLMs are paving the way for improvements in healthcare, particularly in image analysis.

Published in: ASDP 61st Annual Meeting

Publisher: The American Society of Dermatopathology
Date of Conference: November 4-10, 2024