Comparison of neural network architectures for cervical classification

Large-scale screening programmes are operating to reduce the incidence and mortality rate of cervical cancer, a disease which is preventable if detected at the pre-cancerous stage. Screening is based upon the manual inspection of Papanicolaou smears. This is a highly demanding and labour-intensive task and for over thirty years there has been considerable interest in automating the process. The authors are investigating the use of various neural network architectures for the analysis and classification of smear scenes. A feature space was derived from the magnitude of the Fourier transform using a wedge-ring arrangement. The features obtained were invariant to translation and rotation. Neural nets were then used to both reduce dimensionality and to perform the classification. An expertly verified database containing over 2000 high resolution cell images was used to measure the performance of the nets. The single-layer perceptron, multi-layer perceptrons and the constructive algorithm of Fahlman and Lebiere were used as classifiers. The effect of feature extraction nets for pre-processing the feature space was also investigated. Performances were compared in terms of speed, network size and ability to learn and generalise. In addition, classification by a parametric Bayesian classifier allowed comparison with a statistical method. Good classification results were obtained.