Region of interest finding in reduced resolution colour imagery. Application to cancer cell detection in cell overlaps and clusters

This research assesses image analysis and pattern recognition methods that eliminate the image segmentation step normally required before parameter measurement and object classification for use in cancer cell detection in imagery with cell overlaps and clusters. A region of interest finding method previously developed for cancer cell detection in low resolution 3 colour “Pap” smear imagery containing isolated cells has been retrained and assessed on imagery containing computer generated cell overlaps and clumps in addition to isolated cells. The approach taken involves representation of key characteristics of objects to be recognised (or rejected) in terms of complex receptor fields and a flexible decision logic structure suited to high speed implementation. The ultimate goal of this work is to develop methods that work equally well on problems such as cervical cytology where cells tend to be isolated or in clusters and on problems such as tissue analysis where cells are arranged in complex structures where segmenting individual cells accurately is often not possible