Traditional anatomical resources like cadaveric dissection and professional atlases are the gold standard, but face constraints in cost and accessibility. The rapid integration of generative artificial intelligence (AI) in medical education offers new possibilities for custom anatomical illustration. Text-to-image AI generators provide an inexpensive and swift alternative for medical illustration. However, the anatomical fidelity and educational reliability of these models remain insufficiently validated. This study aimed to evaluate the accuracy of four prominent AI image generators in producing anatomically correct illustrations of the human heart, kidney, and brain. Four AI platforms-Copilot, Microsoft Designer, Gemini, and Freepik-were prompted to generate high-resolution, textbook-style illustrations of the heart (sternocostal view), kidney (coronal cross-section), and brain (lateral view). A total of 12 selected images were evaluated by a radiologist and two anatomists utilizing a structured 5-point Likert scale across five domains: textural realism, topological validity, morphometric accuracy, perspective and depth perception, and educational utility. Significant variability in performance was observed across platforms. Gemini consistently demonstrated superior performance, achieving perfect intraclass correlation coefficients (ICCs) and the highest ratings on complex topological tasks, including the representation of coronary vasculature and cortical sulci/gyri. Copilot showed high spatial accuracy in renal hilum organization. While all platforms showed high reliability and accuracy across broad morphological shapes (aortic arch, cerebral poles), accuracy declined sharply as anatomical complexity increased. AI-generated anatomical illustrations currently vary from medically indistinguishable to biologically impossible. While Gemini and Copilot show promise for supplemental educational use, the high frequency of hallucinated anatomical details in other platforms poses a risk for medical misinformation. The study underscores the necessity for expert supervision and for developing AI models trained on curated, peer-reviewed medical datasets.
Albadawi et al. (Sat,) studied this question.