Improving the conformal metric: Geodesic deformable models for medical image analysis

The use of medical image analysis in the, clinical setting has introduced many quantitative techniques for diagnosis, treatment planning, surgical guidance and monitoring of disease processes. A critical component in these applications is the segmentation of the relevant anatomical or physiological regions from the image data. This data might come from a wide array of imaging technologies, many capable of making high resolution, three-dimensional (3D) measurements. Robust segmentation methods can greatly improve the presentation of relevant information, assisting accurate and timely data interpretation. A useful class of segmentation methods often used in medical image analysis is the deformable model. Deformable models incorporate prior information about the shape and location of the objects of interest and have been proven effective in many applications.; One such method, the geodesic deformable model (GDM), has been shown to be useful in applications where topological flexibility is necessary. The GDM is based on an area minimizing gradient flow using a conformal (non-Euclidean) metric. The metric is derived from the Euclidean metric and a data dependent multiplier, referred to as the conformal transformation or speed function. Starting with an initial surface, an ordered family of surfaces is produced that iteratively move, in the steepest descent direction, toward a minimum of an area functional (a geodesic surface). The key to applying the GDM in medical image analysis is the specification of the conformal transformation.; This work is concerned with the application of the GDM in medical image analysis, focusing of the definition of the conformal transformation using the image data and prior information. The desired properties of and restrictions on the transformation are examined in detail, followed by a discussion of the current definition's limitations. A new 3D, multi-scale approach to defining the conformal transformation called the high-confidence boundary integration (HCBI) method is then given with example applications in magnetic resonance imaging (MRI) and x-ray computed tomography (CT). A further refinement of the HCBI method is developed in two dimensions and applied to two medical image segmentation tasks. Finally, the HCBI GDM is used in a clinical application known as virtual colonoscopy, which requires segmentation of the colon lumen from 3D CT datasets. A small study shows that the HCBI GDM is an improvement over the previous GDM implementations in accurately segmenting the colon lumen surface.