Estimation of atlas-based segmentation outcome: Leveraging information from unsegmented images

Segmentation via atlas registration is a common technique in medical image analysis. Devising estimates of such segmentation outcome has been of interest in cases with multiple atlases, both for single-atlas selection and for multi-atlas fusion. This paper studies the estimation of expected Dice's similarity metric for registering atlas-target pairs, by employing registration loops with models of such metric (error) accumulation over these loops. In this framework, the use of registration information also from unsegmented images is proposed and is shown to outperform using segmented atlas images alone. We demonstrate a fast, memory-efficient implementation and single-atlas selection results using a CT and an MR dataset.