Computational modeling of thoracic and abdominal anatomy using spatial relationships for image segmentation

This paper presents an original hierarchical segmentation approach of several thoracic and abdominal structures in CT and emission PET images. Segmentation results will be used to initialize a non-linear registration procedure between these complementary imaging modalities. Therefore, structures involved in the segmentation system must be visible in both CT and emission PET images in order to compute a spatial transformation between them. Thus, the chosen structures include lungs, kidneys and liver (skin and skeleton are also segmented as support structures). In the hierarchical segmentation procedure, the extraction of a given structure is driven by information derived from a simpler one. This information is composed of spatial constraints inferred from the previously segmented structures and expressed by means of Regions Of Interest (ROI) in which the search for new structures will take place. The segmentation of each structure follows a two-phase process: a first stage is composed of automatic thresholding and other low-level operations in the ROI defined by previously segmented objects; a second stage employs a 3D deformable model to refine and regularize results provided by the former step. Visual inspection by medical experts has stated that the proposed segmentation approach provides results which are accurate enough to guide a subsequent non-linear registration procedure.