Automated Cortical Thickness Index Measurement in Pelvic Radiographs Using a Multi-Stage Deep Learning Framework

The Cortical Thickness Index (CTI) measured on hip radiographs is a practical indicator of femoral bone quality and may assist osteoporosis screening and preoperative orthopedic planning. However, manual CTI measurement is time-consuming and subject to inter-observer variability. We developed and validated a multi-stage deep learning framework for automated CTI measurement on routine pelvic radiographs. In this retrospective study, 843 radiographs from a tertiary center were collected, of which 647 met the eligibility criteria for model development and internal testing. An external dataset of 55 radiographs was used for validation. The framework combined RetinaNet-based detection to localize the femur and key landmarks with U-Net-based segmentation to delineate cortical boundaries. To improve anatomical consistency across patients, the measurement level was standardized relative to femoral head diameter. On the internal test set of 129 radiographs, femur detection achieved 96.99% precision and 100% recall. Automated CTI values closely matched expert measurements (0.376 ± 0.087 vs 0.373 ± 0.097), with a mean difference of 0.003, a strong correlation (r = 0.941, p < 0.01), and excellent agreement (ICC = 0.935). Bland-Altman analysis showed narrow 95% limits of agreement (-0.062 to 0.068). Clinically, this pipeline not only reduces reliance on tedious manual measurements but dramatically enhances inter-evaluator reproducibility (ICC = 0.702). In the external cohort (n = 55), the end-to-end pipeline success rate was 83.41%. These findings support the technical feasibility of automated and standardized CTI measurement from pelvic radiographs. Its potential use in osteoporosis screening or preoperative assessment requires independent clinical validation.