Automating segmentation of dual-echo MR head data

Abstract

Multiecho MR acquisition yields various information about tissue and csf characteristics. The analysis of two-dimensional scatterplots generated from dual-echo MR data turns out to be a useful tool. It allows the optimization of MR acquisition parameters for later specific recognition or segmentation tasks. Multivariate statistical classification techniques are applied to dual-echo MR data to segment volume head data into anatomical objects and tissue categories (brain white and gray matter, ventricular system, outer csf space, bone structure, tumor).

To overcome the sensitivity of voxel-based classification to noise we applied a preprocessing technique based on anisotropic diffusion. This preprocessing increases the separability of clusters. We illustrate the robustness of supervised classification with the segmentation of a series of MR head data in a research study.

For a given set of MR parameters we show that the configuration of clusters in feature space is comparable between studies. This allows us to develop an automated clustering technique that considers a priori knowledge about cluster attributes and their configuration in feature space. The automated classification technique omits subjective criteria in the training stage of supervised classification and yields reproducible segmentation results. A new study will check the reliability of the automated classification in comparison with supervised classification.

The segmentation of head MR data in routine clinical applications gives some important qualitative and quantitative information about brain atrophy, brain volume, volume of csf spaces and morphological changes of local brain areas. Combined 3D views of multiple anatomical objects are shown. They clarify the perception of 3D relationship and highlight the locations and types of structural abnormalities.