From text to tissue: Artificial Intelligence‐generation of biological images

Images are essential teaching tools in health professional education. In dental and medical education, for example, high-quality, detailed images are needed to teach gross and micro-anatomy, and cellular mechanisms and compare healthy versus diseased states. Textbook images, created by skilled professional illustrators, are excellent resources but are insufficient to meet the evolving needs of students and educators.1 Freely available online images are not always accurate and reliable.2 Furthermore, educators often face challenges in accessing high-quality, accurate illustrations, providing the angle, level of detail, and magnification to meet specific teaching needs. Which text-to-image generators are suitable for generating biological images? Can the images be improved by altering the prompts? How do different image generators respond to the same prompt? A tooth section showing microanatomical structures. An illustration of a cell membrane representing the fluid mosaic model. A realistic open human mouth showing teeth, tongue, palate, and oral mucosa The anatomical structure of the temporomandibular joint. Five AI image generators were tested with a simple prompt to evaluate their ability to create biological images. Two image generators were excluded for creating unrealistic images (Table 1). A detailed prompt was developed to generate an image of a longitudinally sectioned tooth. The prompt was iteratively improved through an image generator (Copilot) to evaluate its impact on the resulting image (Figure 1). Three AI-image generators (Firefly, Copilot, and Gemini) were prompted with the exact text to compare their performance in creating biological images (Figure 2). Generating appropriate biological images using image generators is iterative and specific to each image generator. What works for Gemini may not work for Copilot. Many image generators like Copilot offer built-in image editors to make minor edits to AI-generated images. One image generator will not produce the same image every time it is prompted with the same text. This feature reduces the reliability of these image generators but also increases the possibility of improving the result. AI-generated images with minor flaws can be used as educational tools by asking students to formatively assess their knowledge by identifying and fixing mistakes in each image. The authors declare no conflict of interest.