Research On Bone Surface Detection Based On 3D CT Images

In modern surgery, to avoid injury of other parts of patients’ body, accurate positioning and segmentation of the target organization or tissue in CT images is usually required. Among all the parts in humans’ body, segmentation of bones is important and difficult. In common, surgeons need to segment the CT images manually, which is definitely a huge work. The main content of this research is algorithm of bone segmentation in 3D CT images, aiming to use computer to segment bones in CT images with accuracy and set surgeons free from the heavy work.In this paper, a combination method of region-based and surface-based is proposed. First, get an initial surface with thresholding method and approach in statistics. Then, for every voxel in the initial surface, calculate its normal direction. Finally, construct a 1-D signal with intensity of voxels along the normal direction and centered at the initial surface voxel and detect the surface to correct the initial surface and get the final bone surface.To get an initial surface containing only target bones, optimal thresholding method is used to do the initial segmentation of CT images in this paper. Then, with 3D flood fill processing, bone voxel set and non-bone voxel set are generated. After this, this paper uses morphology method to remove the non-goal bone voxels from bone voxel set, leaving goal bone voxels alone. However, morphology method may sacrifice the accuracy of segmentation result. To ensure the accuracy of initial surface, an adaptive thresholding approach is used to find best threshold of image blocks through Bayesian decision theory. In this way, the result can be corrected and the initial surface is obtained.Then, this paper proposes an estimation-correction approach for normal direction in 3D images which can improve the precision of normal direction. First, estimate normal directions with first derivative. Then, correct the normal directions through a surface-tracing based method with geometrical features of bone surface and get better results of normal directions. Finally, construct 1-D signals and detect the surface. Furthermore, several parameters need to be set manually in this processing, to ensure the automation and accuracy of this algorithm, this paper proposes an automatically estimation approach for these parameters.Finally, regarding manually segmentation results of experts as golden standards, experiments results based on clinical 3D hip joint CT images are given and compared with several advanced methods, including quantitative measurement results and samples of image results. It is shown that algorithm in this paper performs better than other several methods in clinical 3D hip joint CT images, which can generate closed and accurate surfaces of bones.