Schematic representation of Mandora prediction of joint damages. a, The objective of this study is to automatically detect joint space narrowing and bone erosion damages based on radiographic images in rheumatoid arthritis. The ground truth is the SvH scores manually labeled by human scorers. Our machine learning approach, Mandora, locates joint positions and quantifies joint-specific damage levels with segmentations of joint space regions. b, We use neural network and tree-based regression models to learn contextual information of images as well as the symmetrical interconnections among joints in rheumatoid arthritis.