Explicit segmentation is required for many forms of quantitative neuroanatomic analysis. However, manual methods are time‐consuming and subject to errors in both accuracy and reproducibility (precision). A 3‐D model‐based segmentation method is presented in this paper for the completely automatic identification and delineation of gross anatomical structures of the human brain based on their appearance in magnetic resonance images (MRI).

The approach depends on a general, iterative, hierarchical non‐linear registration procedure and a 3‐D digital model of human brain anatomy that contains both volumetric intensity‐based data and a geometric atlas. Here, the traditional segmentation strategy is inverted: instead of matching geometric contours from and idealized atlas directly to the MRI data, segmentation is achieved by identifying the non‐linear spatial transformation that best maps corresponding intensity‐based features between a model image and a new MRI brain volume. When completed, atlas contours defined on the model image are mapped through the same transformation to segment and label individual structures in the new data set.

Using manually segmented sturcture boundaries for comparison, measures of volumetric difference and volumetric overlap were less than 2% and better than 97% for realistic brain phantom data, and less than 10% and better than 85%, respectively, for human MRI data. This compares favorably to intra‐observer variability estimates of 4.9% and 87%, respectively. The procedure performs well, is objective and its implementation robust. The procedure requires no manual intervention, and is thus applicable to studies of large numbers of subjects. The general method for non‐linear image matching is also useful for non‐linear mapping of brain data sets into stereotaxic space if the target volume is already in stereotaxic space. © 1995 Wiley‐Liss, Inc.