The amount of data needed to describe both volume and surface of 3D objects is often huge and produces bottlenecks at evety step of analysis. Thus, extracting relevant information in this case demands heavy and complex processing techniques. A preprocessing phase is dramatically required : data must be synthesized in order to enhance analysis procedure by providing context-dependent measurements. Parameters set should not be rigid, as it can be radically different from one application to another. The method we propose in this PhD. thesis consists in mapping any digitized 3D solidtaking into account its inner points - into Ellipsoidal Skeleton (or E-skeleton). Based on binaty shape decomposition into a union of simple sub-shapes paradigm, it also gathers relevant information about the geometty and any other set of values that seems interesting, depending on the study context. Each sub-shape and its parameters set are more generically viewed as a feature, which is assumed to be non-decomposable at a given sfmtiJUic lecd. This semantic zoom capability for object description permits a hierarchical approach, i.e. a scale of vision control. Low semantic zoom allows crude approximation for fast pre-classification while high semantic zoom highlights finer details for precise comparison. From this preprocessing stage, tasks such as object recognition and object analysis are made easy and intuitive to perform via features comparison. Any bottleneck is removed, ensuring fast data processing. At last, the versatile structure of the E-skeleton allows not only future improvement but also practical ways to add features to the E-skeleton repertoire. We will show how the E-skeleton is well-suited for medical imaging and how it can be seamlessly integrated into the analysis and diagnostic process.
