Automated bone segmentation in computed tomography using deep learning with distance maps
Abstract: 3D-models of bone structures from computed tomography (CT) data can be used for surgical planning, education and a wide range of other purposes. They can also be used in both digital and 3D-printed format. The currently used process to obtain such models consists of a combination of thresholding, morphological operations and manual adjustments. This can be very time-consuming. Deep Learning (DL) can be used to automatically segment organs such as bones in CT data, and can thus be used to automate such a process. Emerging data also suggests that including distance maps (DM) of ground truth segmentation masks when using DL for segmentation problems can yield improved performance. In this thesis project, a well-known neural network architecture for segmentation was modified in three different ways to include DM during training and prediction. The three modifications were inspired by three types of methods for DM-inclusion used in previous work, but simplified. Estimates of the generalization performances of the three modifications and the network in its original state were compared using both an in-house dataset and a publicly available dataset. The results showed that at least one of the modified networks outperformed the network in its original state in all the tested cases. This indicates that DL-methods for performing bone structure segmentation in CT data could benefit from an inclusion of DM during training and prediction. This was especially indicated when using a multi-task network to perform both segmentation and DM-regression in parallel. However, these results have to be further validated.
AT THIS PAGE YOU CAN DOWNLOAD THE WHOLE ESSAY. (follow the link to the next page)