GPT-5 Series for Dermoscopic Image Labeling

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Generative Artificial Intelligence (AI) models, like ChatGPT, are gaining traction across various fields.
• Dermoscopic image labeling is crucial for accurate skin cancer diagnosis.
• Evaluating the performance of advanced AI models in medical applications is essential.

Purpose of the Study:

• To assess the performance of OpenAI's GPT-5 series (GPT-5, GPT-5.2, GPT-5.4) in dermoscopic image labeling.
• To determine if newer versions of GPT-5 offer improved accuracy in identifying skin conditions, specifically melanoma.
• To compare the effectiveness of different testing formats for AI-driven image analysis.

Main Methods:

• Utilized three distinct testing formats: free-form answers, label selection, and label selection with patient metadata.
• Employed paid ChatGPT Plus subscriptions, disabling the "Improve the model for everyone" option for unbiased results.
• Tested GPT-5 models using the "Thinking" variant across various timeframes for comprehensive evaluation.

Main Results:

• GPT-5.4 demonstrated notable improvements in distinguishing melanoma compared to earlier versions.
• Performance variations were observed across the different testing formats.
• The inclusion of patient metadata alongside images influenced labeling accuracy.

Conclusions:

• Iterative updates to GPT-5 models show enhanced capabilities in dermoscopic image analysis.
• The study highlights the potential of advanced AI in improving diagnostic accuracy for skin conditions like melanoma.
• Further research is warranted to optimize AI model performance and integration into clinical workflows.