Evaluating and improving label fusion in atlas-based segmentation using the surface distance

Abstract

Atlas-based segmentation is an increasingly popular method of automatically computing a segmentation. In the past, results of atlas-based segmentation have been evaluated using a volume overlap measure such as the Dice or Jaccard coefficients. However, in the first part of this paper we will argue and show that volume overlap measures are insensitive to local deviations. As a result, a segmentation that is judged to be of good quality when using such a measure may have large local deviations that may be problematic in clinical practice. In this paper, two versions of the surface distance are proposed as an alternative measure to evaluate the results of atlas-based segmentation, as they give more local information and therefore allow the detection of large local deviations. In most current atlas-based segmentation methods, the results of multiple atlases are combined to a single segmentation in a process called 'label fusion'. In a label fusion process it is important that segmentations with a high quality can be distinguished from those with a low quality. In the second part of the paper we will use the surface distance as a similarity measure during label fusion. We will present a modified version of the previously proposed SIMPLE algorithm, which selects propagated atlas segmentations based on their similarity with a preliminary estimate of the ground truth segmentation. The SIMPLE algorithm previously used the Dice coefficient as a similarity measure and in this paper we demonstrate that, using the spatial distance map instead, the results of atlas-based segmentation significantly improve.