Assessing GPT-4 Multimodal Performance in Radiological Image Analysis

Abstract

AbstractThe integration of Artificial Intelligence (AI) in radiology presents opportunities to enhance diagnostic processes. As of recently, OpenAI’s multimodal GPT-4 can analyze both images and textual data (GPT-4V). This study evaluates GPT-4V’s performance in interpreting radiological images across a variety of modalities, anatomical regions, and pathologies. Fifty-two anonymized diagnostic images were analyzed using GPT-4V, and the results were compared with board-certified radiologists interpretations. GPT-4V correctly recognized the imaging modality in all cases. The model’s performance in identifying pathologies and anatomical regions was inconsistent and varied between modalities and anatomical regions. Overall accuracy for anatomical region identification was 69.2% (36/52), ranging from 0% (0/16) in US images to 100% (15/15, 21/21) in X-ray and CT images. The model correctly identified pathologies in 30.5% of cases (11/36), ranging from 0% (0/9) in US images to 66.7% (8/12) for X-rays. The findings of this study indicate that despite its potential, multimodal GPT-4 is not yet a reliable tool for radiological images interpretation. Our study provides a baseline for future improvements in multimodal LLMs and highlights the importance of continued development to achieve reliability in radiology.