Guiding automatic segmentation with multiple manual segmentations

Abstract

Most image segmentation algorithms are designed to estimate a single segmentation for each image, where the gold standard segmentation is often labeled by a human expert. However, it is common that multiple manual segmentations are available for some images, e.g. independently labeled by different experts. For efficient usages of manual segmentations, we propose to simultaneously produce automatic estimations for each expert. The key advantage of this proposal is that it allows to incorporate the correlations between different experts to improve the accuracy of automatic segmentation. In a brain image segmentation problem, where for each image six manual segmentations are available, we show that jointly estimating several manual segmentations produces significant improvement over independently estimating each of them.