Evaluation of Interactive Segmentation Algorithms Using Densely Sampled Correct Interactions

Abstract

The accuracy and reproducibility of semiautomatic interactive segmentation algorithms are typically evaluated using only a small number of human observers which only considers a very small number of the possible correct interactions that an observer might provide. A correct interaction is one that provides contextual information that would be expected to result in a correct segmentation. In this paper, we demonstrate new evaluation methods for semiautomatic interactive segmentation algorithms that employ simulated observer models constructed from a large number of segmentations computed by uniformly sampling the entire set of possible correct interactions. The advantages of this method are that it is free of observer biases and the large number of segmentations produced for each object of interest to be segmented allow a range of statistical methods to be brought to bear on the analysis of segmentation algorithm performance. The methods are demonstrated using a semi-automated segmentation algorithm for ovarian follicles in ultrasonographic images.