Three-dimensional population-based object modeling for organ boundary definition in low contrast CT images.
Jennifer L. Boes · Deep Blue (University of Michigan) · 1994
Automated organ surface definition in abdominal computed tomography scans (CT) is currently difficult to achieve. The determination of organ location and size is of interest for various medical applications including radiation therapy treatment planning, surgical planning and oncologic monitoring. This object definition domain is of interest due to the availability of three-dimensional data containing objects that vary in non-rigid ways across the population and that consist of low contrast boundaries. A semi-automated method for object definition in this domain is presented. I have developed an object model that accurately represents actual population shape information and applied it to the liver. The model is created by averaging surfaces from a set of normalized liver data sets; the normalization process registers the livers in a standard space using thin-plate spline warping based on a small set of selected landmarks. The model is instantiated by identifying these landmarks in a data set and using them as a basis for model deformation; this preliminary liver representation is fit more closely to the patient data using a combination of the model's Bayesian priors and CT edge information to guide the process. The final output of this fitting technique is a complete object surface adapted to the specific data set. I apply this technique to a set of liver CT scans and demonstrate its effectiveness as a tool for complete liver surface definition, even in low contrast regions. This work illustrates a novel modeling of expected shape populations in a standard space defined by a set of landmarks. The strength of the model fitting technique is its ability to use image-based evidence to fit the model in areas providing strong evidence and to use model-based expected surface locations to define portions of the boundary providing little or no information due to low contrast.