Model-based multiple active contours matching for radiographic images

Medical images are noisy and complex. Segmentation and labeling of X-ray images represent many difficulties. Active contours have become an attractive subject in computer vision. Connectivity and closure properties of these contours help to overcome some important difficulties in computer vision, as edge organization and region merging. Consequently, using deformable contours reduces dramatically the search space dimension. In this paper, we present a model-based approach of multiple dynamic non-parametric curves matching with X-ray images. The model is formed of three parts: (i) image formation, (ii) high-level interaction, and (iii) contours smoothing constraints. The first and second part measure consistency of the reconstructed object with the given image and the relational a priori information of the object, respectively. The scene model represents a hierarchical structure of three processes: lines, regions and relational graphs. An object is modeled as a set of linked subobjects according to a 3-D relational graph which can be projected from a known viewpoint in a 2-D region relational graph. The resultant function is optimized using a descending search method with randomized sampling. Finally, successful results are presented for object matching in semitransparent noisy synthetic scenes