AI-natomy: human anatomy through the eyes of artificial intelligence. Is there a distinction between reality and imagination?

Abstract

BACKGROUND: The application of artificial intelligence (AI) in medicine has been a strong trend in recent years. Many generative tools have a model that processes a text command (so-called "prompt") and generates a response in the form of text (e.g. ChatGPT) or an image (e.g. Midjourney, DALL-E). Due to the wide availability, these tools are often used without proper supervision for educational purposes. The aim of this study was to assess the state-of-the-art use of AI-based generative tools in creating an image-based anatomical atlas using only images created by AI. MATERIALS AND METHODS: The author's provided anatomical prompt into the Midjourney and individual images were made for each of the eight organ systems - digestive, facial, genitourinary, respiratory, skeletal, muscular, circulatory, and neurological. The systems were divided into two sections: a general sketch of body system (BS) and a detailed description of one of their representative structures (DS). The authors selected two of the highest images for each collection, correspondingly. The obtained data was assessed qualitatively depending on the presence of a specific structure, a physiological alteration, or the presence of an additional AI hallucination. Obtained results were compared with The Federative International Programme for Anatomical Terminology - Terminologia Anatomica 2nd version (TA2) and statistically analyzed using the chi-square test of independence (two-sided α = 0.05, β = 0.2) to evaluate the relationship of image complexity and the categories studied. RESULTS: A total of 12126 general BS and 2503 specific DS items according to TA2 were obtained and analyzed. As for the presence of structures, they were generally present in 17.3% (n = 2094), absent in 53.3% (n = 6462) and the generated scenery did not allow a clear assessment of the presence of a structure in 29.4% (n = 3570). Only 7.0% (n = 553) of the current structures were in physiological order. However, 6.0% (n = 471) of the particular items were classed as AI hallucinations. When comparing BS and DS findings, the tool significantly outperformed DS in all categories: structural presence (21.7% BS vs 40.3% DS; p < 0.001), physiological order (4.6% BS vs 15.5% DS; p < 0.001), and AI imagination (6.3% BS vs 4.9% DS; p = 0.034). CONCLUSIONS: The current state-of-the-art of anatomy based on generative tools does not allow for the high-quality application of AI to generate anatomical atlases. Especially, inexperienced users may unknowingly cultivate inconsistent scientific understanding. The results obtained are expected to improve in the future due to the continuous addition of annotations to libraries and the improvement of algorithms, nevertheless more research is required.

Ähnliche Arbeiten