Improving Artificial Intelligence-Based Diagnosis on Pediatric Skin Lesions.

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Artificial intelligence (AI) algorithms for melanoma classification often face limited generalizability due to training data biases.
• Current AI models predominantly use adult-derived datasets, potentially underperforming on pediatric cases.

Purpose of the Study:

• To evaluate the impact of incorporating pediatric images into an AI training dataset on melanoma classification performance.
• To compare AI model performance on adult and pediatric test datasets before and after the addition of pediatric images.

Main Methods:

• Trained two AI models: Model A on adult data (37,662 images) and Model A+P on adult data plus pediatric images (1,536).
• Compared model performance using Area Under the Receiver Operating Characteristic Curve (AUC) on separate adult and pediatric test sets.
• Utilized Gradient-weighted Class Activation Maps (Grad-CAM) and background skin masking to analyze decision-making factors.

Main Results:

• The addition of pediatric images (Model A+P) significantly improved AI performance on pediatric test images.
• Performance on adult test images remained comparable between Model A and Model A+P.
• Background skin features contributed to the improved pediatric performance observed with Model A+P.

Conclusions:

• Incorporating diverse pediatric data enhances the generalizability of AI models for melanoma detection in dermatology.
• Carefully curated and labeled datasets from varied populations are crucial for developing robust AI diagnostic tools.
• AI models can be improved for broader clinical application by including pediatric dermoscopic images.