Applying Design of Experiment to Optimise Artificial Neural Network for Classification of Cervical Cancer

Classification of cervical cancer with high accuracy remains as one of the most challenging issues in the automated computer aided diagnosis system. In this study, the significance of four selected features, i.e. nucleus area, cytoplasm area, nucleus mean intensity and nucleus to cytoplasm ratio (N/C ratio) are investigated. These features are extracted from the public image database of Herlev University Hospital, Denmark. Classification of cervical cell is formulated into a two-class problem, dividing the obtained 917 cervical cell images into normal and abnormal cases. T-test analyses show that the four specified features have good potentials as the artificial neural network (ANN) classification elements. The success rates of different ANN framework for classification achieve an average of 86%. The best optimised ANN framework (i.e. eight hidden neurons, traingd training algorithm and learning rate of 0.09) achieves an average success rate of 98.50%.