Discriminative dimensionality reduction for patch-based label fusion

In this last decade, multiple-atlas segmentation (MAS) has emerged as a promising technique for medical image segmentation. In MAS, a novel target image is segmented by fusing the label maps of a set of annotated images (or atlases), after spatial normalization. Weighted voting is a well-known label fusion strategy consisting of computing each target label as a weighted average of the atlas labels in a local neighborhood. The weights, denoting the local anatomical similarity of the candidate atlases, are often approximated using image-patch similarity measurements. Such an approach, known as patch-based label fusion (PBLF), may fail to discriminate the anatomically relevant patches in challenging regions with high label variability. In order to overcome this limitation we propose a supervised method that embeds the original image patches onto a space that emphasizes the appearance characteristics that are critical for a correct labeling, while supressing the irrelevant ones. We show that PBLF using the embedded patches compares favourably with state-of-the-art methods in brain MR image segmentation experiments.