LIGHT MICROSCOPIC HETEROGENEITY OF HUMAN GLIOBLASTOMAS INVESTIGATED WITH AN UNSUPERVISED NEURAL NETWORK (SOM)

Article type: Research Article

Authors: Iglesias-Rozas, José R. | Grieshaber, Frieder

Affiliations: Institut für Pathologie (Neuropathologie). Katharinenhospital,Stuttgart | Institut für Biomedizinische Technik, Stuttgart, Germany

Note: [] Address for correspondence: Priv.-Doz. Dr. med. J.R. Iglesias-Rozas Klinikum Stuttgart. Katharinenhospital Institut für Pathologie (Neuropathologie) (Ärzt. Dir.: Prof. Dr. med. A. Bosse) Kriegsbergstr. 60 D-70174 Stuttgart Tel: ++ 49 +711/278 4918 Fax: ++49 +711/278 4909 E-Mail: [email protected] [email protected]

Abstract: The histological variability of glioblastomas precludes the modern assimilation of theses tumors into a single histological tumor group. There is evidence for the genetic variability of glioblastomas. As an alternative to statistical evaluation we investigated 1266 human glioblastomas in order to discover whether they can be correct classified using SOM (Self-Organizing Maps). In all tumors 45 histological features including age and sex of the patients, were examined. The description of the presence of a specific histological feature is given on a scale of four class. No prior statistical knowledge or clustering is needed. Five clusters of glioblastomas with a maximum significance were found. Cluster C1 contains glioblastomas with a great component of glioblasts and astroblasts. Cluster C2 includes 93.75% of all gliosarcomas. Cluster C3 contains 80.28 % of all monomorphe glioblastomas. Cluster C4 shows similarities with the features of relative poverty of vessels, scanty vessels anomalies and little thrombosis. Cluster C5 contains 60.68% of all giant cell glioblastomas. Placing a series of component windows with their maps side by side allows the immediate investigation of the dependencies on variables. The nets SOM allow in our study a realistic histological classification, comparable to the actual classification made by WHO, as well as the visualization of multidimensional histological features of human glioblastomas. With SOM one can learn to discriminate, discard and delete data, select meaningful histological and clinic variables and consecutively to influence the results of patients' management.

Keywords: neural networks, SOM, glioblastoma, clustering, classification, brain tumors, glioma

Journal: Electronic Journal of Pathology and Histology, vol. 6, no. 4, pp. 02-02, 2000