P2‐096: Fast and Robust Segmentation of Brain Magnetic Resonance Images

Quantitative and objective analysis of brain MRI images typically requires segmentation of images. Multi-atlas segmentation (MAS) has been shown to perform well in segmenting sub-cortical structures from images. The major problem of this approach is that it is computationally expensive limiting its use in clinical practice. We study if computation speed could be improved to meet clinical requirements without compromising the accuracy considerably. In MAS, several atlases are registered non-rigidly to patient data and propagated segmentations are fused using, e.g., classifier fusion. In this work, we perform the registration via a separate template as an inter-mediate step between atlas- and patient spaces. The mappings between the atlas and template spaces can be pre-computed and only one registration from the patient space to the template space must be performed during segmentation. Segmentation accuracy was evaluated using two databases: ADNI (www.loni.ucla.edu/ADNI, 60 cases) and IBSR (http://www.cma.mgh.harvard.edu/ibsr, 18 cases). The Dice similarity index was used to compare automatically and manually generated segmentations (0 = no overlap, 1 = perfect overlap). The similarity index for the hippocampus was 0.885 (ADNI) and 0.825 (IBSR) when the original MAS was used. When a separate template (MNI template, http://neuro.debian.net/pkg/fsl-atlases.html) was applied, the corresponding values were 0.861 and 0.800. Although the accuracy is decreased by 0.02, the value 0.861 is still equal to the accuracy obtained for two manual segmentations in other studies using ADNI. Using IBSR data the computation time of the template based approach was 2 minutes on a standard laptop which was only 9 % of the original MAS. The computation time includes also segmentation of other regions (sub-cortical structures, tissues and parcellations). The average decrease in the similarity index was 0.015 using IBSR data. Computation times of several hours have been reported for MAS which makes its use in clinical practice difficult. Our implementation allows running MAS in about 20 minutes on a standard PC. However, the use of a separate template reduced the time to 2 minutes without compromising the accuracy drastically. We consider that this makes the use of this method attractive as a routine tool in clinical practice.