The role of atlases and multi-atlases in brain tissue segmentation based on multispectral magnetic resonance image data

Atlas assisted image segmentation has been quite popular in medical imaging during the last two decades. The atlas is able to provide prior information on the imaged organ’s shape, appearance, and local texture or intensity distribution. In case of segmenting images via pixelwise classification, the final segmentation result is obtained through a fusion of the classification outcome with the local atlas information. In other words, the atlas guides the classifier towards the shape of local structures normally situated at the given location. This paper proposes to demonstrate the advantages a multi-atlas can bring in a segmentation process of the main tissues in infant brain based on multi-spectral MRI records. Three supervised machine learning methods are deployed to segment brain tissues, with and without the use of the atlas. Differences are evaluated using statistical accuracy indicators. Atlases improved the overall segmentation accuracy by 2.5-3.5%, depending on the deployed classifier method.