- G Wolf: Department of Pathology, Charité Medical School, Humboldt University, Berlin, Germany.
The quantification of chromatin structures is an important part of nuclear grading of malignant and premalignant lesions. In order to achieve high accuracy, computerized image analysis systems have been applied in this process. Chromatin texture analysis of cell nuclei requires a suitable texture model. A hierarchic model seemed to be most compatible for this purpose. It assumes that texture consists of homogeneous regions (textons). Based on this model, two approaches to texture segmentation and feature extraction were investigated using sections of cervical tissue. We examined the reproducibility of the measurement under changing optical conditions. The coefficients of variations of the texture features ranged from 2.1% to 16.9%. The features were tested for their discriminating capability in a pilot study including 30 cases of cervical dysplasia and carcinoma. The overall classification accuracy reached 65%. This study presents an automated technique for texture analysis that is similar to human perception.