Label fusion in multi-atlas based segmentation with user-defined local weights

Multi-atlas based segmentation is a popular method to automatically segment a target image, in which the correspondence to already segmented atlas images is used to construct multiple segmentations for a single structure in the target image. These multiple segmentations are then combined into a single segmentation for the target image in a process called label fusion. In the past, the result of multi-atlas based segmentation has mostly been evaluated using a volume overlap measure. However, such a measure can only be used to assess the global quality of a segmentation and does not take into account local differences in for example the clinical relevance of a certain region of the segmentation. We propose to use voxel-based weights in the evaluation of segmentations and show that by using these weights already during the label fusion process, one is able to obtain multi-atlas based segmentation results with an improved clinical relevance compared to unweighted atlas based segmentation. A method is proposed to implement this for multi-atlas based segmentation of the prostate.