Architecture of the deep neural network for segmenting joint space regions and predicting damage scores. This neural network contains an encoder to extract information at multiple scales, a decoder to decode the abstracted information from feature maps, and a regressor to quantify the damage score of a joint. Multiple convolution, max-pooling, up-convolution layers and one dense layer are used. The decoder generates the segmentation mask, which is further used as input for the regressor. The encoder, the decoder and the regressor are connected through concatenation layers.