What is Segmentation? Explain how it works, and provide images as examples of the technique.

Where different structures have similar radiodensity, it can become impossible to separate them simply by adjusting volume rendering parameters. The solution is called segmentation, a manual or automatic procedure that can remove the unwanted structures from the image

Segmentation can be performed manually or (semi)automatically. Segmentation algorithms are often based on the principle of region growing. Placing one or more seed points initiates the segmentation of the target structure. From these seed points, more and more neighboring voxels that fulfill predefined criteria are included in the segmentation (14). The technique can be applied in two ways: segmentation of the desired tissue or segmentation of the undesired tissue with subsequent removal from the data. The latter method removes only interfering tissue (bone or densely enhanced veins) from the CT angiography data and retains soft tissue as well as contrast-enhanced vessels for further evaluation. To refine the boundary of the segmented structures, morphologic dilation operations may be applied. A particular problem in threshold-based segmentation algorithms are areas with close contact of two tissue types with comparable attenuation, such as bone and contrast-enhanced vessels (course of the ICA through the skull base; intraforaminal sections of the vertebral artery) (Fig 6). Although the process of segmentation is semiautomatic, user interaction is necessary to set additional seeding points or to intervene in cases of inclusion of neighboring structures due to leakage of the region-growing algorithm. These procedures can be time-consuming and may exceed practical limits in routine clinical work flow.

The segmentation involves two issues; separating the bone from the surrounding soft tissue and separating different parts of the bone from each other. The goal with the segmentation process is a surface model representing the anatomy of a specific patient that can be incorporated into the simulator system and used for practice or pre-operative planning.
The method that has been employed for the automatic segmentation process is the Morphon method [1-3]. This is a general non-rigid registration method that has been developed at our department. The method takes a prototype image/volume and deforms it until it fits the target image/volume. The registration process is an iterative algorithm where each repetition passes through the following steps: * Displacement estimation * Deformation field accumulation and regularisation * Prototype deformation The algorithm is initiated on a coarse resolution scale to catch large, global displacements, and continues to finer resolution scales until an optimal match between the two datasets is obtained. For a more detailed description of this method the reader is referred to references [1-3].
As mentioned above the algorithm works on both 2D and 3D datasets. In this project we work with 3D CT volumes collected from patients with this type of fracture. The prototype is a volume where each voxel is labeled such that it belongs either to the pelvis, the femur or the background. By deforming this segmented representation of the hip area until it fits the corresponding structure in a specific patient we obtain a volume with a labeled representation of that specific patient's anatomy. From the labeled volume it is easy to generate surface models of the pelvis and femur as two separate objects.