Detection of malignancy associated changes in cervical cells using statistical and evolutionary computation techniques

Abstract Malignancy Associated Changes are subtle alterations in the morphology and nuclear texture of cells in the vicinity of a malignant lesion. The phenomenon was first described in 1959, and has been the subject of considerable research in the four intervening decades, due to its potential utility to cancer screening programs. In this thesis the history of research into malignancy associated changes is reviewed, and the major findings of previous workers summarized. Original work aimed at improving the accuracy of classification of Pap smear slides is described in detail. A novel algorithm, which incorporates a genetic algorithm for feature selection and training of a neural network, is described. The algorithm was tested upon a large artificial dataset consisting of points from nested spheres in multiple dimensions. It was able to select the most discriminatory features and classify data with 99% accuracy on 80% of runs for two dimensional data, and on 90% of runs for three-dimensional data. The algorithm was also tested on two real data sets from the UCI Machine Learning Repository, the “sonar” data and the “ionosphere” data. On both of these datasets the algorithm produced a classifier using a subset of features which performed as well as previously reported classifiers using the full feature set. This algorithm was then tested on a large dataset of cell images, and its performance compared with that of the standard stepwise linear discriminant analysis approach. Both of these approaches produced similar results, which are comparable to those of previous workers in this field. Interestingly, runs of the genetic algorithm with different random number seeds tended to select different feature subsets, which produced approximately equivalent performance. This finding indicates that amongst the features used, which were selected from those previously identified in the literature as useful for MACs detection, many subsets exist which are equally discriminatory.