Pattern recognition based segmentation method of cell nuclei in tissue section analysis

A pattern recognition-based segmentation (PRS) system was developed for segmenting cell nuclei in tissue sections of urine bladder tumors and brain tumors. One hundred and thirty eight image samples from HE-stained tissue sections of urine carcinoma and brain astrocytomas were selected. Half of them were used for designing the PRS-system and the rest for evaluating its performance. The PRS-system can be designed by employing one of three classifiers: minimum distance, Bayesian, and multilayer perceptron (MLP) classifier. Classifier design was based upon two sets of training data, the nuclei set and the surrounding tissue set, which were derived from textural features. According to the pathologists´ evaluation, 88% of the segmented nuclei were registered correctly and 12% incorrectly. The MLP classifier proved superior in sensitivity and discrimination of nuclei from background when compared with Bayesian and minimum distance classification performance. The PRS-system proved efficient when tested under different types of histopathological tissue samples providing an index for potential generalization of the technique.