Deep learning approach to creating bone models

The paper discusses the usefulness of deep neural networks in tasks of automatic segmentation of bone structures from CT data. Three different neural network models (PSPNet, U-net and Segnet) were trained to perform the segmentation task. Pixels in the CT image are assigned to 3 following classes: background, patient outline and bones. The mean and classwise Intersection over Union (IoU) measure was evaluated for each network outcome. The most exact segmentation results were obtained with the use of U-net model, with mean IoU value equal to 93.2%.