Comparison of traditional brain segmentation tools with 3D self-organizing maps
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 self-organizing 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 S 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.
Unable to display preview. Download preview PDF.
- 1.Chalana, V., Kim, Y.: A methodology of Image Segmentation Algorithms on Medical Images. In (M.H. Loew and K.M. Hanson, eds.) SPIE MT96 2710-20 (1996) 178–189.Google Scholar
- 2.Dean, D., Buckley, P., Bookstein, F., Kamath, J., Kwon, D., Friedman, L., and Lys, C.: Three Dimensional MR-based Morphometric Comparison of Schizophrenic and Normal Cerebral Ventricles. In (K.-H. Höhne and R. Kikinis, eds.) Visualization in Biomedical Computing '96. Lecture Notes in Computer Science. 1131 (1996) 363–372.Google Scholar
- 3.Joliot, M., Mazoyer, B.M.: Three-dimensional segmentation and interpolation of magnetic resonance brain images. IEEE Trans. Med. Imaging 12 (1993) 269–277.Google Scholar
- 4.Kohonen, T.: Physiological interpretation of the self-organizing map. Neural Networks 6 (1993) 895–905.Google Scholar