In multi-atlas based segmentation propagation, segmentations from multiple atlases are propagated to the target image and combined to produce the segmentation result. Local weighted voting (LWV) method is a classifier fusion method which combines the propagated atlases weighted by local image similarity. We demonstrate that the segmentation accuracy using LWV improves as the number of atlases increases. Under this context, we show that introducing diversity in addition to image similarity by using least-angle regression (LAR) criteria is a more efficient way to rank and select atlases. The accuracy of multi-atlas segmentation converges faster when the atlases are selected in the order of LAR. We test the method on a hippocampal atlas set of 138 normal control (NC) subjects and another set of 99 Alzheimer's disease patients provided by ADNI. The result shows that LAR selection is more efficient than similarity based atlas selection. Fewer atlases were required using LAR selected atlases to achieve the same accuracy as fusing atlases from image similarity based selection.