According to one embodiment there is provided a method of selecting a plurality of M atlases from among a larger group of N candidate atlases to form a multi-atlas data set to be used for computer automated segmentation of novel image data sets to mark objects of interest therein. A set of candidate atlases is used containing a reference image data set and segmentation data. Each of the candidate atlases is segmented against the others in a leave-one-out strategy, in which the candidate atlases are used as training data for each other. For each candidate atlas in turn, the following is carried out: registering; segmenting; computing an overlap; computing a value of the similarity measure for each of the registrations; and obtaining a set of regression parameters by performing a regression with the similarity measure being the independent variable and the overlap being the dependent variable.