Computer Graphics and Visualization • Publications

This thesis investigates methods for medical visualisation. A common approach to visualisation of 3D data volumes from CT or MRI systems is to extract surfaces that describe the shape of anatomical objects. Such surfaces can help a clinician with diagnosis, planning, navigation, manipulation, visualisation and surgical simulation. For these purposes two requirements are important: accuracy and quality. Accuracy describes the conformance of a mesh to an object surface, whereas quality relates to the compactness and well-formedness of the mesh, e.g., well-shaped triangles with a high area to perimeter ratio are important for subsequent processing steps, for example finite element analysis. Two methods were developed that meet both requirements: the first is an iso-surface based SurfaceNets method, the second is a scale-space edge-detection method. The basis of both methods is formed by the application of alternating mesh improvement and extraction techniques. Mesh improvement is performed by variants of Laplacian smoothing and edge swapping. The meshes created by the SurfaceNets method are equal in accuracy and superior in triangle quality compared to the Marching Cubes iso-surface extraction algorithm. The scale-space edge-detection method provides robust edge-detection and localisation in combination with a set of mesh improvement tools that improve the quality and prevent geometric and topological errors from occurring (e.g., holes and inverted mesh elements). For each surface extraction technique the robustness and effectiveness of the alternating approach is shown. Another method that is presented in this thesis is a new interactive segmentation method. Conventionally, manual segmentation is performed by outlining a desired shape in a stack of 2D images. Our new method circumvents the problems related to contour connecting by operating directly in 3D. Starting with a template object that consists of three orthogonal, planar, and connected contours, a desired shape can be created by moving points of each contour. The user can add new contours in three orthogonal directions; the new contours are automatically linked at the intersections of the current model with a slicing plane. The linked contours reduce the amount of tedious manual interaction, and this is reduced even further by the addition of an automatic edge-detection method. It is shown that the fast interactive segmentation method provides effective feedback and control for semi-automatic segmentation methods. Finally, a statistical shape analysis of two carpal bones is performed. Active shape models are initialised by a large number of example shapes. From this set modi of variation are calculated that describe shared characteristics of the input shapes. Effectively, shapes are parametrised by a linear combination of each modus of variation. Conventionally, an explicit landmark correspondence is required between the shapes. In our approach this correspondence is determined automatically. By doing so, the initial shapes are allowed to be defined in a free-form manner. The fast interactive segmentation method proves to be a valuable tool for creating these initial shapes. In conclusion, the set of described methods together provide the means to create 3D surface representations from 3D medical image data in a short amount of time,even if the data is noisy. The final result is an accurate and high-quality mesh that is well suited for further modelling and post-processing steps, such as visualisation, mechanical analysis, and simulation.

P.W. de Bruin, Accurate and high-quality surface extraction from medical image data, PhD Thesis, Delft University of Technology, 2004.