Limitations of Artificial Intelligence Generated Images for Hand Surgery Patient Education

Area of Science:

• Medical imaging
• Artificial intelligence in healthcare
• Surgical education

Background:

• Artificial intelligence (AI) is rapidly advancing, with significant potential to enhance clinical workflows and patient understanding in medicine.
• Text-to-image AI generators are becoming increasingly sophisticated, raising questions about their utility in specialized fields like surgery.

Purpose of the Study:

• To evaluate the anatomical accuracy of common hand surgery procedure images generated by popular AI text-to-image models.
• To determine if AI-generated images are suitable for patient education materials in hand surgery.

Main Methods:

• Five AI text-to-image generators (Craiyon, DALL-E, DeepSeek, Gemini, Midjourney, Stable Diffusion) were prompted to create labeled images of surgical approaches for conditions including carpal tunnel syndrome and Dupuytren contracture.
• Generated images were assessed for legibility, detail, clarity, anatomical realism, accuracy, appropriate surgical site, and absence of fabricated anatomy, using a 10-point control score as a benchmark.

Main Results:

• All AI generators performed significantly below the control, with 99.8% of images containing fabricated anatomy.
• DALL-E showed the highest overall scores, while Craiyon scored the lowest.
• While some AI models (DALL-E, DeepSeek, Gemini, Midjourney) demonstrated comparable detail and clarity to the control, all significantly underperformed in anatomical realism and accuracy.

Conclusions:

• Current AI text-to-image generators, despite producing visually appealing images, frequently fail to generate anatomically correct representations for hand surgery.
• Significant advancements in AI model training and fine-tuning are required before these tools can be reliably used for creating accurate patient education materials in surgical contexts.