Assessing deep learning accuracy in the measurement of radiographic parameters in pediatric hip X-rays

Assessing radiographic parameters in pediatric pelvic X-rays is crucial for evaluating hip development, yet existing deep learning (DL)-based methods lack both age-specific reliability analysis and a comprehensive solution for measuring multiple key parameters.

This retrospective study developed and validated a DL-based system using separate, nonoverlapping datasets of 1495 and 1300 anteroposterior (AP) pelvic radiographs of normal Korean children for model training and evaluation, respectively. The system measured the acetabular index (AcI), Shenton line (ShL), pelvic rotation index (PRI), and pelvic tilt index (PTI). Subgroup analyses were conducted to evaluate the effects of age-related pelvic bone development. Evaluation metrics included the intraclass correlation coefficient (ICC), Pearson correlation coefficient (r), mean absolute error (MAE), root mean square error (RMSE), Hausdorff distance (HD), and Frechet distance (FD). Agreement between the system's and clinician's measurements was assessed using Bland-Altman analysis.

For all evaluation data, automatically measured AcI, PRI, PTI, and ShL values strongly matched and correlated with radiologist-assessed values (AcI: ICC = 0.89, r = 0.91, MAE = 2.07°, RMSE = 2.99°; PRI: ICC = 0.94, r = 0.94, MAE = 0.03, RMSE = 0.04; PTI: ICC = 0.97, r = 0.97, MAE = 0.04, RMSE = 0.09; ShL: HD = 3.62 mm, FD = 2.27 mm). The subgroup analysis revealed that the system's performance varied with age-related differences in pelvic bone development.

The DL-based system exhibited high reliability and accuracy in measuring radiographic parameters for differentiating normal from dislocated hips and assessing pelvic radiograph quality.