The use of subject-specific FE models in clinical practice requires a high level of automation, validation and accurate evaluation of analysis prediction capabilities. This study presents a novel high-order finite element method (p-FE) for generating FE models based on CT data. The geometry is represented by smooth surfaces accurately and an internal smooth surface is used to separate the cortical and trabecular regions upon which a p-FE auto-mesh is constructed within each region (cortical or trabecular). QCT Hounsfield Unit (HU) values are recalculated using a moving average method regardless of the FE mesh and isotropic inhomogeneous linear material properties are assigned to the FE models by a spatial function computed from the QCT data. Anisotropic material properties are also considered in the FE model and their influence on the mechanical response is examined. To validate the FE results we performed QCT scans on three fresh frozen proximal femurs (30 years old male, 20 and 54 years old female) followed by mechanical in-vitro experiments at different inclination angles, measuring head deflection and strains at several points with a total of 77 experimental values recorded. The FE results, strains and displacements were compared to the in-vitro experiments. The linear regression of the experiment results compared to the FE predictions demonstrates the high accuracy and reliability of our methods (slope of 0.955 and R^2=0.957). These results are significantly better than any previous published results known to the authors.