Predicting biological sex in pediatric skeleton X-rays using artificial intelligence

Abstract

Artificial intelligence (AI) is increasingly applied in medical imaging, yet its ability to predict biological sex from pediatric radiographs remains unclear. This study investigates the performance of convolutional neural network (CNN) models in sex classification using a large dataset of pediatric trauma imaging and compares results with human raters. Radiographs from computed and digital radiography systems were processed to normalize grayscale and enhance contrast. The EfficientNet family of CNN models (B0-B7) was trained on this dataset, with attention to balancing the test set by age, sex, and fracture visibility. A subset of 1,000 images was independently assessed by human raters for comparison. AI models achieved a mean precision of 0.731 ± 0.035, recall of 0.718 ± 0.110, accuracy of 0.722 ± 0.032, and F1-score of 0.724 ± 0.050 across all network variants. Performance improved with age, peaking in the 13-18 group. Pelvic X-rays achieved the highest classification metrics. Human raters showed significantly lower agreement. AI can classify biological sex from pediatric radiographs with high accuracy, surpassing human performance. Results vary by age and body region, supporting the potential for AI-assisted imaging in pediatric clinical practice.

Ähnliche Arbeiten