Magnetic resonance imaging (MRI) of bones has added value for fracture risk assessment in osteoporosis, a disease of weak bones. However, manual segmentation of bone images is time-intensive, causing slow throughput for test results and inefficient risk assessment for patients. In this work, we implemented an automatic proximal femur segmentation algorithm by modeling a convolutional neural network (CNN) as a pixel-wise binary classification. The accuracy of automatic segmentation was investigated by analyzing similarity between automatic and manual ground-truth segmentation. In addition, we compared the time required for manual fine-tuning of the CNN segmentation with original manual segmentation.