Patient-specific atlas is a key technology for the recognition of the human anatomy from 3D medical images. Automated recognition of the pancreas is one main issue for computer-assisted diagnosis and therapy systems in the abdomen because many diseases of the pancreas are not accompanied by noticeable symptoms. In patient-specific atlas generation, hierarchical and mosaicing methods have been proposed to cope with individual differences in the position, orientation, and shape of the pancreas. Even though segmentation accuracy was improved by these methods, it remains lower than for other abdominal organs, such as the liver and the kidneys. The location of the pancreas strongly correlates with the location of vasculature systems, especially the splenic vein. In this paper, we propose a new structure specific atlas generation method that considers the structural information in atlas generation. As for the structural information, we enhance the vasculature using a vesselness filter. Similar volumes in a training dataset with respect to the vasculature structure are selected and used for atlas generation. Using 150 cases of contrast-enhanced 3D abdominal CT volumes, our experiment improved the mis-segmentation of the surrounding organs or such soft tissues as the duodenum.
展开▼
