Comparison of traditional brain segmentation tools with 3D self-organizing maps

Abstract

Algorithm-assisted 3D MR brain segmentation may be significantly faster than manual methods and produce visually pleasing results. We tested two- and three-dimensional region growing (2DRG and 3DRG) and selforganizing map (SOM) algorithms for segmentation of the cerebral ventricles. The SOM algorithm provides the greatest times savings, 12:1, over manual segmentation. Concern for reproducibility of algorithm-assisted segmentation motivated an intra-operator comparative study of these and manual segmentation methods. One of us, DK, segmented the cerebral ventricles from 5 3D MR-scan data sets three times manually and with the three algorithms. When variability is measured as the shape variance of derived landmarks sets, the three algorithm-assisted methods show less intra-operator variability than manual segmentation. The 2DRG and 3DRG segmentations show more variability than SOM. Of the 4 methods, SOM segmentation requires the fewest operator decisions.