Can AI-Based ChatGPT Models Accurately Analyze Hand-Wrist Radiographs? A Comparative Study

Area of Science:

• Orthodontics and Dental Anthropology
• Artificial Intelligence in Medicine
• Radiographic Analysis

Background:

• Accurate bone age assessment is crucial for orthodontic treatment planning.
• Conventional methods and deep learning models require significant infrastructure and training.
• Large language models (LLMs) offer a potential infrastructure-independent alternative.

Purpose of the Study:

• To evaluate the effectiveness of LLM-based chatbot systems (ChatGPT) in predicting bone age and growth stages.
• To explore LLMs as practical, infrastructure-independent alternatives to current methods.
• To compare LLM performance against conventional methods and convolutional neural network (CNN) models.

Main Methods:

• Three ChatGPT models (GPT-4o, GPT-o4-mini-high, GPT-o1-pro) analyzed 90 anonymized hand-wrist radiographs.
• Radiographs represented pre-peak, peak, and post-peak growth stages, with equal sex distribution.
• Expert orthodontists established reference standards using Fishman's SMI system and Greulich-Pyle Atlas.

Main Results:

• All LLMs demonstrated significant agreement (p < 0.001) in bone age prediction; GPT-o1-pro showed highest concordance (r=0.546).
• GPT-o4-mini-high achieved 72.2% accuracy within a ±2 year deviation for bone age.
• GPT-4o exhibited the highest agreement (κ=0.283, p < 0.001) for growth stage classification.

Conclusions:

• General-purpose LLMs can assist in predicting bone age and growth stages, each with unique strengths.
• LLMs offer contextual reasoning and preliminary assessment capabilities without domain-specific training.
• Further development is needed, but LLMs show potential as supportive tools in orthodontics, complementing clinical examination.