Classification of astrocytomas and malignant astrocytomas by principal components analysis and a neural net.

The classification of astrocytomas, astrocytomas with anaplastic foci and glioblastoma multiformes is not always straightforward because the tumors form a histological continuum. The use of principal component analysis (PCA) and neural nets in the classification of these tumors is explored. PCA was performed on 14 histological features recorded from 52 gliomas classified by the Radiation Therapy Oncology Group method (17 astrocytomas, 18 astrocytomas with anaplastic foci, 17 glioblastoma multiformes). Four of the 14 possible 'scores' derived from this analysis were selected to summarize the histological variability seen in all the tumors. These scores were mostly significantly different between tumor types and were thus used to successfully train a neural net to correctly classify these tumors. The first principal component (score) supported the use of increasing cellularity, mitoses, endothelial proliferation, and necrosis in differentiating between the tumor categories, but accounted for only 39% of the variability seen. Other histological features that were significant components of the other scores included the presence of multinucleated or giant cells, gemistocytes, atypical mitoses and changes in nuclear chromatin. Computer programs derived from the methodology described provide a way of standardizing glioma diagnosis and may be extended to assist with management decisions.