Classifier selection strategies for label fusion using large atlas databases

Structural segmentations of brain MRI can be generated by propagating manually labelled atlas images from a repository to a query subject and combining them. This method has been shown to be robust, consistent and increasingly accurate with increasing numbers of classifiers. It outperforms standard atlas-based segmentation but suffers, however, from problems of scale when the number of atlases is large. For a large repository and a particular query subject, using a selection strategy to identify good classifiers is one way to address problems of scale. This work presents and compares different classifier selection strategies which are applied to a group of 275 subjects with manually labelled brain MR images. We approximate an upper limit for the accuracy or overlap that can be achieved for a particular structure in a given subject and compare this with the accuracy obtained using classifier selection. The accuracy of different classifier selection strategies are also rated against the distribution of overlaps generated by random groups of classifiers.