Evaluating General Vision-Language Models for Clinical Medicine

Recently emerging large multimodal models (LMMs) utilize various types of data modalities, including text and visual inputs to generate outputs. The incorporation of LMMs into clinical medicine presents unique challenges, including accuracy, reliability, and clinical relevance. Here, we explore clinical applications of GPT-4V, an LMM that has been proposed for use in medicine, in gastroenterology, radiology, dermatology, and United States Medical Licensing Examination (USMLE) test questions. We used standardized robust datasets with thousands of endoscopy images, chest x-ray, and skin lesions to benchmark GPT-4Vs ability to predict diagnoses. To assess bias, we also explored GPT-4Vs ability to determine Fitzpatrick skin tones with dermatology images. We found that GPT-4V is limited in performance across all four domains, resulting in decreased performance compared to previously published baseline models. The macro-average precision, recall, and F1-score for gastroenterology were 11.2%, 9.1% and 6.8% respectively. For radiology, the best performing task of identifying cardiomegaly had precision, recall, and F1-score of 28%, 94%, and 43% respectively. In dermatology, GPT-4V had an overall top-1 and top-3 diagnostic accuracy of 6.2% and 21% respectively. There was a significant accuracy drop when predicting images of darker skin tones (p<0.001). GPT-4V accurately identified Fitzpatrick skin tones for 56.5% of images. For the multiple-choice-styled USMLE image-based test questions, GPT-4V had an accuracy of 59%. Our findings demonstrate that the current version of GPT-4V is limited in its diagnostic abilities across multiple image-based medical specialties. Future work should be done to explore LMMs sensitivity to prompting as well as hybrid models that can combine LMMs capabilities with other robust models.