Transfer Functions Design Based on Statistical Shape Models

This presentation examines statistical models for transfer functions based on an initially generated set of manually assigned transfer functions with respect to a very specific type of data set and a strictly delimited type of application. The process of transfer function design is decoupled from the specialized knowledge about the transfer function domain (intensity, gradient magnitude etc.).Transfer function design is difficult because of the high degrees of freedom and the lack of a truly goal-directed process. Existing approaches have been developed for automatic of semi-automatic transfer function design and can be categorized in image-driven and data-driven techniques. To concentrate on the anatomical or functional structures which are interesting for the user an application-driven method is needed. For a well-defined application scenario it is possible to reduce the complexity of transfer function generation by restricting the classification process to structures of interest for a specific examination procedure. At first, transfer functions are manually generated for an initial collection of volume data sets that has been recorded for one specific clinical purpose. A single transfer function is represented by a set of parameters of geometric primitives (ramps or trapezoids). Each of these individually assigned transfer functions can be regarded as a point sample in the (high-dimensional) parameter space of the transfer function model. From this set of point samples in parameter space a statistical shape model is created by applying Principal Component Analysis. A higher-level transfer function models, with only a very limited set of parameters, based on this analysis is established to make the process of transfer function setup very simple and intuitive.