TY - JOUR TI - Improved automatic bone segmentation using large-scale simulated ultrasound data to segment real ultrasound bone surface data DO - https://doi.org/doi:10.7282/t3-z81z-td09 PY - 2020 AB - Automatic segmentation of bone surfaces from ultrasound images is of great interest in the ultrasound-guided computer assisted orthopedic surgery field. These automatic segmentations help the system locate where the bone surface is in the image which can allow for proper surgical manipulation. Methods that involve using image processing tools have previously been used to perform the segmentations however, they have faced problems due to the noise and various imaging artifacts associated with ultrasound data. Most recently, methods based on deep learning have achieved promising results. However, a drawback is that these methods require large number of training dataset. Therefore, new methods which can overcome these drawbacks need to be investigated in order to accurately segment bone surfaces from real ultrasound data. This thesis introduces the concept of training the deep learning methods with large-scale simulated bone ultrasound data and investigating how using large-scale simulated data along with limited real ultrasound data affects the segmentation performance of the deep learning network. A transfer learning approach and using a training dataset consisting of both real and simulated ultrasound bone surface data was applied for the investigation. We show that by using simulated bone ultrasound data, the success of traditional deep learning methods increases compared to using small-scale real ultrasound data only. Data used in the study consisted of real ultrasound data collected from different subjects and utilizing 3D Slicer and PLUS for generating simulate ultrasound data. Various networks were trained in order to determine how well the network of a certain dataset is able to perform automatic segmentations on the same type of data. Additionally, networks trained with both large-scale simulated US data and limited real ultrasound data were trained and tested on real ultrasound data to determine if using large-scale simulated data improves network performance. The automatic segmentations of the neural networks were compared against manual segmentations of the same data by calculating the Sorensen- Dice Coefficient and Average Euclidean Distance. Results of the thesis show that using large- scale simulated ultrasound data can be used to train a neural network to segment real ultrasound data if both types of datasets are used together to develop the network. KW - Ultrasound segmentation KW - Skeleton -- Imaging KW - Biomedical Engineering LA - English ER -