Scoliosis is a medical condition where one’s spine curves sideways. As is commonly observed in children, scoliosis needs diagnosing early to avoid deterioration. Conventionally, doctors spend much time studying the back X-ray of the patients. Modern methods use deep learning to process the X-ray. However, the radiation given out by X-ray machines, which hurts children’s health, cannot be avoided by these methods. We present a new workflow to diagnose scoliosis by synthesizing the X-ray from the RGB-D images of the back of the patients using deep learning models, which can rid the whole process of radiation. We used anatomical landmarks to reveal shape of the spine. Firstly, landmark detection on RGB-D images using the High-Resolution Nets was conducted. Then we used the pix2pix model to synthesize X-ray from RGB-D images and landmarks. Detailed experiments were conducted to tune the models. Our synthetic images were clear and close to the ground truth. Only 520 training samples were available before the deadline. However, good performance on this small dataset could prove our idea. If more data is available, X-ray machines are no longer needed for diagnosis. Our project has a bright prospect.