SoftSTAPLE: Truth and performance-level estimation from probabilistic segmentations
Neil I. Weisenfeld, Simon Keith Warfield · 2011
We introduce here a new algorithm, called softSTAPLE, for computing estimates of segmentation generator performance and a reference standard segmentation from a collection of probabilistic segmentations of an image. These tasks have previously been investigated for segmentations with discrete label values, but few techniques exploit the information available in probabilistic segmentations. Our new method may be used to evaluate classification algorithms, to fuse “weak” classifiers in a performance-weighted fashion, or to combine the results of a previous fusion of manual segmentations in an hierarchical manner. We describe and validate our new algorithm, and compare its performance to other techniques in two applications with “real-world” data.