Automatic measurement of proximal femoral morphological parameters using point cloud semantic segmentation technology

This paper proposes a method for the automatic measurement of proximal femoral morphological parameters based on CT images. First, construct a statistical model of the femur with generalization properties and perform Mask labeling. Second, automatically segment the femur model from CT images to obtain the femur sample to be tested. Then, use the GBCPD point cloud registration algorithm to complete a fast hierarchical registration between the two models, establishing point-to-point correspondences. Based on these correspondences and femoral morphological features, automatically segment the test femur sample into the shaft, neck, and head. Numerical methods are used to determine the femoral shaft axis (using PCA combined with least-squares cylinder fitting), the femoral head center and radius (least-squares sphere fitting), as well as the eccentricity and femur length. We conducted reproducibility tests of this method on 213 femurs and compared the results between automatic segmentation/measurement and manual segmentation/measurement. The Dice similarity coefficients for the femoral head, neck, and shaft reached 0.98, 0.95, and 0.99, respectively. The reproducibility errors of the anatomical standards (angles, dimensions) for automatic measurement were all lower than the errors between manual measurements, indicating that the parameter values obtained by this method exhibit good consistency with the corresponding parameter values manually identified by medical experts in the original CT images. This effectively minimizes subjective influences and can assist orthopedic surgeons in large-scale measurement analysis.