Dall-E in hand surgery: Exploring the utility of ChatGPT image generation

Abstract

Background

Artificial intelligence (AI) has significantly influenced various medical fields, including plastic surgery. Large language model (LLM) chatbots such as ChatGPT and text-to-image tools like Dall-E and GPT-4o are gaining broader adoption. This study explores the capabilities and limitations of these tools in hand surgery, focusing on their application in patient and medical education.

Methods

Utilizing Google Trends data, common search terms were identified and queried on ChatGPT-4.5 and ChatGPT-3.5 from the following categories: “Hand Anatomy”, “Hand Fracture”, “Hand Joint Injury”, “Hand Tumor”, and “Hand Dislocation”. Responses were graded on a 1–5 scale for accuracy and evaluated using the Flesch-Kincaid Grade Level, Patient Education Materials Assessment Tool (PEMAT), and DISCERN instrument. GPT 4o, DALL-E 3, and DALL-E 2 illustrated visual representations of selected ChatGPT responses in each category, which were further evaluated.

Results

ChatGPT-4.5 achieved a DISCERN overall score of 3.80 ± 0.23. Its responses averaged 91.67 ± 0.29 for PEMAT understandability and 54.67 ± 0.55 for actionability. Accuracy was 4.47 ± 0.52, with a Flesch-Kincaid Grade Level of 9.26 ± 1.04. ChatGPT-4.5 consistently outperformed ChatGPT-3.5 across all evaluation metrics. For text-to-image generation, GPT-4o produced more accurate visuals compared to DALL-E 3 and DALL-E 2.

Conclusions

This study highlights the strengths and limitations of ChatGPT-4.5 and GPT-4o in hand surgery education. While combining accurate text generation with image creation shows promise, these AI tools still need further refinement before widespread clinical adoption.