Brain atlases have played an important role in multi-atlas based brain image segmentation. It has been reported that top-ranked atlas selection with mutual information (MI) similarity can result in a better segmentation performance compared to that from a random set of atlases. However, the mutual information based image similarity between atlas and target image is not directly related to the final segmentation performance, especially when small sets of atlases are selected ¹. In this study, we test different atlas pre-selection strategies to evaluate the brain segmentation accuracy and efficiency, which is based on the multiple granularity analysis. In M1 pipeline, the lateral ventricle (LV) of target image is primarily labeled before the multi-atlas based segmentation. By adopting the same pipeline for segmentation, results based on the proposed pre-selection method were compared to that using MI similarity, random selected atlases, and 4 labels, which contain the gray matter, the white matter, the lateral ventricle, and the surrounding CSF space. This study shows that the proposed method can reduce heavy computational burden in atlas pre-selection process, while remains high accurate multi-atlas based brain segmentation.The atlas database in current study was selected from JHU T1 Geriatric Multi-Atlas Inventory. The atlas database contains T1 images of brain from 77 subjects (age=34±25.00), consisting three age groups: geriatric (age=71.3±7.0), adult (age=35.6±11.4), and pediatric (age=10.8±2.5). Those 77 brain images were manually parcellated into 286 structures defined in the JHU brain atlas. We proposed one-label approach using the lateral ventricle (LV) and the 4 labels (4L) -based, which is related to the gray matter, the white matter, the lateral ventricle, and the surrounding CSF space. These are adopted to calculate the Dice between the target image and each atlas. As a measure of similarity, Dice is a statistical validation metric to evaluate the overlapping accuracy of paired structural regions. It is commonly used to evaluate the agreement of segmentations between an automated segmentation estimate and the manual one. In addition, we tested MI-based approach using voxel intensity information from the whole brain image. In brief, each target image is segmented by total 24 atlas subsets, including the randomly selected and the top 5, 10 selected based on 4L, LV, and MI-based methods for geriatric, adult and pediatric populations. Randomly selected subsets with 5, 10 atlases were also generated for each target image for comparison. In addition, pre-selection efficiency was also evaluated and compared by the computation time on an Intel Pentium CPU.The segmentation accuracy was evaluated by Dice between the automatic method and the manual annotation. Figure 1 illustrate the segmentation accuracy with different pre-selected strategies for different ages. After the atlas ranking based on one of the three similarity measures, 5 and 10 atlases were selected for the multi-atlas segmentation. The 4L and LV approaches constantly showed higher accuracy. Significant difference was found among the four measures for the geriatric and adult populations. The 4L, LV and MI-based pre-selection methods show obviously different performance in accuracy. In addition, atlas ranking by Dice of 4L and paired LV label would cost 20s, while it could be 400s by the MI of paired T1 images (Table 1).Adopting the same segmentation pipeline, we compared the proposed pre-selection methods to MI-based pre-selection in terms of accuracy and efficiency. The quantitative analyses based on DICE overlap showed the improved accuracy by all three methods, while the 4L and LV-based approaches consistently led to higher level of accuracy at a given number of employed atlases, comparable to MI-based pre-selection. It also can be observed that with increasing number of selected atlas subsets, the segmentation accuracy would also increase in the 4L-, LV-and MI-based atlas pre-selection methods. By paired t-test, the DICE analyses with different age groups showed that significant statistical differences were found for the proposed methods compared to MI and random pre-selection for the geriatric and adult populations. All these comparisons demonstrate a comparable atlas pre-selection with high performance for all methods. From the above presented results, the proposed 4L and LV-based atlas pre-selection shows a promising potential in increasing accuracy. For the time efficiency comparison, 4L- and LV-based pre-selection demonstrates a much higher speed than the MI-based one. The computation for MI of a paired brain T1 image can be 20 times of the proposed pre-selection measures. All these results imply that 4L- and LV-based atlas pre-selection is a reliable method in terms of both accuracy and efficiency.