Application of self-organizing maps and clustering to the segmentation of 3D and 4D medical images

P. Hobelsberger. Application of self-organizing maps and clustering to the segmentation of 3D and 4D medical images. 2002.

  • Petra Hobelsberger
Abstract Segmentation of 3D data sets from computed tomography is important for the visualization of the different structures contained in these images and for further processing of the data. If several 3D images of a moving object are taken at different time steps, a fourth dimension is added to the data. Due to the large amount of data in these 4D data sets, processing them is a fairly time consuming task. Computed tomography systems and, therefore, the applications which handle their data are commonly used for medical diagnosis. This thesis introduces a hierarchical approach to the problem of volume segmentation which uses self-organizing maps in combination with different clustering algorithms. The method is based on the generic approach described in [DROBICS M. et al.: Mining Clusters and Corresponding Interpretable Descriptions - A Three-Stage Approach. Expert Systems, Volume 19, July, 2002] and is applied to the problem of finding the heart muscle—more precisely the left ventricle—in a series of 3D CT images taken from a beating human heart.The process of training a three-dimensional self-organizing map with data from a single 3D image and the subsequent classification of its nodes with clustering and region growing methods is described. Preprocessing was introduced to additionally increase computation speed. Several preprocessing methods such as downsampling are mentioned. It is shown that a rough segmentation of the data can be achieved using this approach.The main advantage of this method is that the segmentation of a single 3D image can be propagated to the subsequent time steps. This propagation is made possible by the abstraction introduced by representing the 3D image by a self-organizing map. This feature of the proposed approach can also be used to simplify user interaction, since every editing performed on the segmentation of one 3D data set can be applied to all other time steps automatically. The proposed method was implemented in a demo system which was used to compute first results. These results seem promising but have to be further improved in order to be applicable in real life situations. A summary of methods which might be able to improve computation speed and segmentation quality is included in this work to simplify further development and research.