Due to the weak boundary, narrow or even disappeared joint space and varying topology in challenging CT, accurate segmentation of the femur from hip joint is still a difficult task. To address this problem, the proposed method combines anatomical information of relative location of bone tissues and neighboring slices to predict a statistical model for Joint space identification. A novel idea in this paper for separating the femur head from the rest parts is the usage of barrier line, and the characteristic of improved harmonic field make the barrier line smoother and attracts the barrier to boundary of femur more closely, which guarantees the accuracy in segmenting the bone tissues with narrow space. We have evaluated our method on 40 hips including pathologies. The experimental results demonstrate the effectiveness and accuracy of the proposed approach.