Evaluating GPT-5 for Melanoma Detection Using Dermoscopic Images.

Area of Science:

• Artificial Intelligence
• Dermatology
• Medical Imaging

Background:

• Melanoma is a deadly skin cancer where early detection is critical for survival.
• Artificial intelligence (AI), specifically large language models (LLMs), offers potential for improving early melanoma detection.
• Systematic assessment of LLMs, such as GPT-5, for melanoma detection is lacking.

Purpose of the Study:

• To evaluate the diagnostic performance of GPT-5 on dermoscopic images for melanoma detection.
• To compare GPT-5's capabilities against previous GPT-4 models in classifying skin lesions.

Main Methods:

• GPT-5 was tested on 600 dermoscopic images from the ISIC Archive and HAM10K datasets.
• The model performed three diagnostic tasks: primary diagnosis, top-3 differential diagnoses, and malignancy discrimination.
• Performance was measured using sensitivity, specificity, accuracy, and F1 score against histopathology-verified ground truth.

Main Results:

• GPT-5 showed modest accuracy for primary diagnosis but significantly improved performance for differential diagnoses (sensitivity >93%, specificity >86%).
• For malignancy discrimination, GPT-5 exhibited more balanced sensitivity and specificity compared to GPT-4 variants.
• GPT-5 outperformed GPT-4 models in differential diagnosis accuracy and overall classification reliability.

Conclusions:

• GPT-5 demonstrates significant potential as a clinical decision support tool in dermatology, surpassing GPT-4 derivatives in differential diagnosis.
• The model's ability to aid medical education in melanoma detection is highlighted.
• Caution is advised due to GPT-5's tendency to misclassify melanoma as benign, necessitating further refinement.