Pelvic MRI-based machine learning models for age estimation and age threshold classification in living children and young adults.

OBJECTIVES: Accurate bone age assessment is crucial in forensic medicine and pediatrics. This study aimed to systematically characterize the MRI developmental patterns of the iliac crest and ischial tuberosity and to develop machine learning models for age estimation and age threshold classification. METHODS: A retrospective analysis was conducted on pelvic MRI scans from 680 individuals aged 6-30 years. A four-stage method was used to assess ossification stages of the iliac crest and ischial tuberosity, followed by a descriptive analysis. We trained and test multiple machine learning models for: (1) regression models for continuous age prediction, and (2) classification models for determining legal age thresholds (12, 14, 16, 18 years). Models were validated on both internal and external test sets. RESULTS: Inter- and intra-observer agreements were good (k > 0.9). the minimum ages for complete fusion of iliac crest and ischial tuberosity were 15.00 years and 15.75 years in males, while both 14.00 years in females. Chronological age and ossification stage showed strong positive correlations (r > 0.8). No significant differences were found regarding sex, sequence, or side (p > 0.05), except the side difference in female ischial tuberosity (p = 0.008). For age regression, the optimal model achieved a mean absolute error of 2.957 years on the internal set and 2.252 years on the external set. For classifying legal age thresholds (12, 14, 16, 18 years), models demonstrated outstanding performance, with the highest AUCs of 0.977, 0.992, 0.969, and 0.931 on the internal set, and 0.955, 0.984, 0.997, and 0.997 on the external set. CONCLUSION: This study provided foundational MRI reference data for pelvic apophyseal development and showed the exciting potential of integrating pelvic MRI with machine learning for age estimation.