The 3D bone shape extracted from the CT image is used for diagnosis, rehabilitation or follow-up studies. However, the necessity of the CT acquisition under multiple postures raises problems of increased radiation exposure and medical costs. In this study, we address this problem by reconstructing CT image from only radiography image. Our proposed method consists of three components: (1) an image-to-image transfer network in order to synthesize DRR (Digitally Reconstructed Radiography) image from real radiography image. (2) a framework in order to build a training dataset for the image-to-image transfer network. (3) a CNN that reconstructs 3D bone shape from DRR image. In this study, we introduce a method for image-to-image transfer and a reconstruction algorithm using the transfer method. I discuss feasibility of the proposed frameworks in order to build training dataset for image-to-image transfer network. And, I evaluated the accuracy of the reconstructed 3D bone shapes from synthesized DRR image generated from radiography image using proposed image-to-image transfer method.