Construction of statistical shape atlases for bone structures based on a two-level framework.

BACKGROUND The statistical shape atlas is a 3D medical image analysis tool that encodes shape variations between populations. However, efficiency, accuracy and finding the correct correspondence are still unsolved issues during the construction of the atlas. METHODS We developed a two-level-based framework that speeds up the registration process while maintaining accuracy of the atlas. We also proposed a semi-automatic strategy to achieve segmentation and registration simultaneously, without knowing any prior information about the shape. RESULTS We have separately constructed the atlas for the femur and spine. The experimental results demonstrate the efficiency and accuracy of our methods. CONCLUSIONS Our two-level framework and semi-automatic strategy are able to efficiently construct the atlas for bone structures without losing accuracy. We can handle either 3D surface data or raw DICOM images.