Automated 3D Bone Segmentation using Deep Learning in Scoliosis

Sammanfattning: Background: Scoliosis is a condition where a person's spine is curved and rotated in three dimensions. In severe cases, surgery is needed to straighten the spine. Before such complex procedures, meticulous planning is needed. By performing a CT-scan, images of the spine can be acquired. From these images, it is possible to segment the spine in three dimensions to facilitate the preparation. However, doing these segmentations manually is very time consuming. Aim: The aim of this thesis was to develop a model for automatic segmentation of spines suffering from scoliosis in CT-images. Depending on the results, an expansion of the model to general bone segmentation was to be evaluated. Methodology: To develop a model, deep learning with the established architecture U-net was used. For training, validation and testing of the networks, 31 datasets were available. The datasets consisted of CT-image stacks covering different bodyparts in different patients. Several models were trained and tested to evaluate the performance of different hyperparameters and segmentation algorithms. An approach for 3D segmentation based on voting between different anatomical planes was compared to a basic 2D segmentation method. Finally, the best model was extented to general bone segmentation. Result: Our best model, Voting 3D (edge), scored an average Dice score of 0.927 (±0.020) and Jaccard score of 0.865 (±0.034) on the scoliosis datasets. The extended network for general bone segmentation scored an average Dice score of 0.938 (±0.052) and Jaccard score of 0.888 (±0.086). Conclusion: The results show that an automatic model based on the U-net can be used to segment spines with scoliosis in CT-images. The results also suggest that by training on more types of bones, a satisfactory model for general bone segmentation can be obtained.