Segmentation of Brain Compartments

Hentschel S.1, Hojjatoleslami A.2, Kruggel F.3, Mittelhäusser G. 4, Palubinskas G. 2, Rajapakse J.C. 5

1 Department of Computer Science, University of Leipzig, 2 Max-Planck-Institute of Cognitive Neuroscience, Leipzig, 3 University of California, Irvine, 4 Department of Computer Science, University of Kaiserslautern, 5 Nanyang Technological University, Singapore

Image segmentation is a basic problem in low-level image processing: extracting information from a dataset without any semantic interpretation. It can be compared to the first level of human vision - perception without interpretation. Segmentation algorithms label a given scene, i.e., address a label referring to a specific class to each voxel. The segmentation problem can be divided into three components: information extraction, classification and regularization. First, the relevant classification feature to be defined and extracted. With classical approaches, the intensity level of a voxel is a classification feature and thus the first problem can be overcome. However, for some applications we have to extract textural features or to combine different information sources. Next, classification separates the set of relevant features into different clusters. Here, the problem is to find the number of clusters and their localization in the feature space. Finally, introducing a priori knowledge may improve classification. Errors arise due to noise, intensity inhomogeneities, and the partial volume effect. Regularization and statistical methods (MRF, Maximum Entropy, etc.) are used to impose constraints on the solution.

The size of head compartments (head and brain volume, intracranial volume, grey and white matter volume, cerebro-spinal fluid volume) and their ratios was determined on the basis of MRI of the head acquired in a reference population of 502 healthy subjects. Age-matched subgroups were selected to reveal gender-related differences and changes with age. Normative data is provided in the form of simple equations that allow transforming measured compartment volumes into z-scores, offering the possibility to relate individual data to a larger population.


Example of the use of z-transformed ratios for the quantitative description of a brain. EX1 (left) has a z-transformed ratio BRV/ICV of +2.64 (corresponding to an unusual large brain volume inside the cranium), EX2 (right) a z-transformed ratio BRV/ICV of -1.77 that is in the transition zone to a brain atrophy.

Hentschel S., Kruggel F. (2004) Segmentation of the Intracranial Compartment: A Registration Approach. In: Jiang T. (ed.), Medical Imaging and Augmented Reality (Beijing), Lecture Notes in Computer Science Vol. 3150, pp. 253-260. Springer, Singapur.

Kruggel F., von Cramon, D.Y. (1999) Alignment of MR Brain Datasets with the Stereotactical Coordinate System. Medical Image Analysis 3, 175-185.

Kruggel F. (2006) MRI-Based Volumetry of Head Compartments: Normative Values of Healthy Adults. NeuroImage 30, 1-11.

Mittelhäusser G., Kruggel F. (1994) Fast Segmentation of Brain Magnetic Resonance Tomograms. In: Ayache N. (ed.), Computer Vision, Virtual Reality and Robotics in Medicine (Nizza), Lecture Notes in Computer Science Vol. 905, pp. 237-241. Springer, Heidelberg.

Palubinskas G., Descombes X., Kruggel F. (1998) An Unsupervised Clustering Method using the Entropy Minimization. In: Caelli T., Maeder A. (eds.), 14th International Conference on Pattern Recognition (Brisbane), pp. 1816-1818. IEEE Computer Society Press, Los Alamitos.

Rajapakse J.C., Kruggel F., von Cramon, D.Y. (1998) Segmentation of MR Images with Intensity Inhomogeneities. Image and Vision Computing 16, 165-180.