Deep learning-based automatic measurement of the femoral head ossification center in healthy Korean children: development of a novel radiographic growth chart - Report - MDSpire
Advertisement
Deep learning-based automatic measurement of the femoral head ossification center in healthy Korean children: development of a novel radiographic growth chart
Automated Deep Learning Assessment of Femoral Head Ossification in Korean Children
Overview
This study developed a deep learning algorithm to automatically measure femoral head ossification center (FHOC) size from pelvic radiographs of healthy Korean children. The method demonstrated accurate segmentation and size measurement, enabling creation of an age- and sex-specific radiographic growth chart for FHOC development.
Background
Pelvic radiography is routinely used to evaluate pediatric hip joints, where the appearance and size of the femoral head ossification center (FHOC) serve as key indicators of skeletal maturity and developmental disorders. Manual measurement of FHOC size is limited by observer variability and time constraints. Advances in artificial intelligence offer potential for automated, standardized assessments to improve reproducibility and efficiency in pediatric musculoskeletal imaging.
Data Highlights
Dataset Subset
Number of Radiographs
Percentage
Training
910
53.4%
Validation
195
11.4%
Test
600
35.2%
The test set was stratified into 24 age- and sex-specific subgroups (12 age groups × 2 sexes), each with approximately 50 bilateral pelvic measurements.
Key Findings
A three-stage cascaded deep learning algorithm was developed for ROI detection, FHOC segmentation, and size measurement from AP pelvic radiographs.
The model was trained and validated on 1705 radiographs from healthy Korean children aged infancy through adolescence.
Automated FHOC size measurement was based on maximum transverse diameter between cortical landmarks detected on segmentation masks.
The test set included balanced age- and sex-specific subgroups to ensure robust performance evaluation.
The algorithm demonstrated accurate and reproducible FHOC segmentation and size measurement, overcoming limitations of manual assessments.
A radiographic growth chart for FHOC size was established to provide standardized reference values for clinical use.
Clinical Implications
Automated FHOC size measurement using deep learning can enhance objectivity and efficiency in pediatric hip assessments, reducing observer variability and workload. The established growth chart offers clinicians standardized reference data to aid in diagnosing and monitoring developmental hip disorders such as skeletal dysplasia and developmental dysplasia of the hip.
Conclusion
This study successfully developed and validated a deep learning-based method for automated FHOC size measurement, enabling creation of a standardized radiographic growth chart. This approach holds promise for improving pediatric hip joint evaluation in clinical practice.
