Pattern recognition techniques in recognition of neoplastic changes in images of cell nuclei

Abstract

A pattern recognition methods were used to classify specimens obtained from exfoliated cell nuclei of urinary bladder. Throughout progress of cancer, neoplastic changes are mainly observed in cell nuclei [1]. The aim of this study was to examine whether no supervised classification based on the k-nearest neighbors rule (k-NN) could give better results of recognition than a multistage classifier constructed and applied previously. At the beginning a computer-assisted system for identification of neoplastic urothelial nuclei was used [2]. The system analyzed Feulgen stained cell nuclei obtained with bladder washing technique and analysis was carried out by means of a digital image processing system designed by the authors. Features describing nuclei population were defined and measured. Then a multistage classifier was used to identify positive and negative cases [3]. Then the same features were used and we used k-NN rule to classify analyzed cases. The analysis was carried on 131cases. This number of data became numerous enough for analyzing k-NN classiffiers for 8 variables. The result for the standard k-NN classifier was: 74% specificity and 75% sensitivity. The result for the parallel k-NN classifier was: 89% specificity and 74% sensitivity. Finally, the hierarchical k-NN classifier was taken under consideration and the obtained result was: 68% specificity and 91% sensitivity. The sensitivity and specificity were calculated for all the methods for two classes: control and malignancy. From the above results we concluded that the hierarchical classifier which gave the best outcome from the sensitivity point of view and it could be used in construction of computer-aided systems dedicated for aid in detection of urinary bladder cancer. The only draw back would be a not as high level of specificity as sensitivity what can generate some number of false positive results.