Early Diagnosis of Scoliosis in Children from RGB-D Images Using Deep Learning

Scoliosis is typically diagnosed from X-ray images, but diagnostic X-rays increase the risk of developmental problems and cancer in those exposed. This project aims to eliminate children’s X-ray exposure in the early diagnosis of scoliosis by generating X-ray images of children’s backs from the corresponding harmless RGB-D images. In this project, a deep learning model based on HRNet was built and trained to detect landmark locations of the backs on the RGB-D images first. With the detected landmarks, the X-ray images of the children’s backs were eventually synthesized from the corresponding RGB-D images by another deep learning model that was built and trained based on the pix2pix model.