Multi-STEPS: Multi-label similarity and truth estimation for propagated segmentations

Quantitative analysis in medical imaging often relies on anatomical segmentation of MR or CT images. Several multi-atlas based segmentation propagation methods have recently been published due to the accurate structural segmentations produced by propagating and combining manual delineations from multiple templates in a database. We propose a new multi-label local ranking strategy for template selection based on the locally normalised cross correlation (LNCC) and an extension to the classical STAPLE algorithm by Warfield et al. [15]. It addresses the known problem of local vs. global image matching and the bias introduced in the performance estimation due to structure size. Results show a significant improvement in terms of segmentation accuracy for key brain areas when compared to state-of-the-art brain parcelation algorithms and reduced discontinuity and fragmentation between structures due to the MRF derived smoothness.