Supervised brain tissue segmentation using a spatially enhanced similarity metric

Abstract

Many medical applications commonly make use of brain magnetic resonance images (MRI) as an information source since they provide a non-invasive view of the head morphology and functionality. Such information is given by the properties of head structures, which are extracted using segmentation techniques. Among them, multi-atlas-based methodologies are the most popular, allowing to consider prior spatial information about the distribution of brain structures. These approaches rely on a nonlinear mapping of the information of the most relevant atlases to a query image. Nevertheless, methodology effectiveness is highly dependent on the mapping function and the atlas relevance criterion, being both of them based on the selection of an MRI similarity metric. Here, a new spatially weighting measure is proposed to enhance the multi-atlas-based segmentation results. The proposal is tested in an MRI segmentation database for state-of-the-art image metrics as means squares, histogram correlation coefficient, normalized mutual information, and neighborhood crosscorrelation and compared against other spatial combination approaches. Achieved results show that our proposal outperforms baseline methods, providing a more suitable atlas selection.