Scoliosis, which can usually be found in teenagers or children, is a troublesome disease. Scoliosis in children is traditionally diagnosed by measuring the curvature of the spine based on X-ray images of patients' backs. However, the X-ray machines expose patients to radiation that can increase the risk of developmental problems and cancer in those exposed. With the development of deep learning in computer vision, convolutional neural network (CNN) based methods have now been adopted in the diagnosis of scoliosis, which have yielded acceptable results. In this project, we proposed to build two deep learning models based on High-Resolution Net (HRNet) and the pix2pix model to perform landmark detection on RGB-D images of the backs and to synthesize X-ray images of the children's backs from the corresponding RGB-D images and detected landmarks, respectively. Children can, in this way, obviate X-ray exposure in the early diagnosis of scoliosis.
The project can be divided into two main phases: landmark detection and X-ray synthesis.
The first objective is to perform landmark detection on RGB-D images with High-Resolution Net (HRNet).
The second objective is to synthesize X-ray images of children's backs from the corresponding RGB-D images and the detected landmarks with the pix2pix model.