Deep learning-based determination of hip geometrical features from X-ray images

Hip morphological disorders could lead to joint wear that, in severe cases, would require a total hip replacement. We propose a deep learning-based approach to determine hip geometrical quantitative features directly from raw X-ray images. Those features include femoral head diameter (HD), intertrochanteric distance (ITD), shaft width (SW), center edge angle (CEA), medial proximal femoral angle (MPFA), and neck shaft angle (NSA). The VGG-16 network was trained and tested using an 853 lower limb radiographs dataset. Results from the test samples showed 4.6, 5.48, 6.5, 16, 7, and 3.3% errors on HD, ITD, SW, CEA, MPFA, and NSA respectively, indicating that the proposed approach performs as well as traditional techniques.