How many templates does it take for a good segmentation?: error analysis in multiatlas segmentation as a function of database size

摘要: This paper proposes a novel formulation to model and analyze the statistical characteristics of some types segmentation problems that are based on combining label maps / templates atlases. Such segmentation-by-example approaches quite powerful their own for several clinical applications they provide prior information, through spatial context, when combined with intensity-based methods. The proposed models class multiatlas as nonparametric regression in high-dimensional space images. presents systematic analysis estimation's convergence behavior (i.e. characterizing error function size database) shows it has specific analytic form involving parameters fundamental problem chosen anatomical structure, imaging modality, registration method, label-fusion algorithm, etc.). We describe how estimate these show brain structures exhibit trends determined analytically. framework also provides per-voxel confidence measures segmentation. large database sizes can be predicted using small-sized databases. Thus, small databases exploited predict required ("how many templates") achieve "good" segmentations having errors lower than specified tolerance. cost-benefit is crucial designing deploying systems.