Introduction

The automation of the cervical cancer screening process has been a long-standing problem in image analysis that has drawn the attention (and frustration) of research groups over the world and over the years.

The main tasks of a potential device are similar to those in almost any other image analysis system. Image capture is followed by the discrete blocks of segmentation, feature extraction and classification.

However, despite the many excellent results in the feature extraction and classification stages, the automatic segmentation of cell images has been uniformly identified as a major hurdle in the development of a working system.

Perhaps the main reason for this, and the inadequate success rates that have been reported in the past, is due to the fact that it is a deceptively difficult problem. Often the images appear trivial to segment and indeed the most basic global thresholding technique will produce good results in some examples, but the Pap smear screening application requires high degrees of accuracy over extremely large data sets and this is a much more difficult problem.

This situation has undermined many attempts to produce an accurate automated cervical cancer screening system and has led to the realisation that the segmentation stage is the key to a working machine [2].