Combined malignancy associated change and contextual analysis for computerized classification of cervical cell monolayers.

Previous efforts to automate cervical cell analysis have focused on algorithms that use measurements of high resolution images of individual cells for 'rare event' detection of abnormal cells, sometimes in combination with cluster analysis of low resolution images. A new methodology was developed in which intermediate cell markers (indicating malignancy associated change) and contextual analysis of slide architecture were performed on the same digital images at 0.33 mu pixel resolution. A total of one hundred and forty-six cases that were either negative or at least high grade squamous intraepithelial lesions (HGSIL) were prepared as monolayers, and about 20 fields per slide were analyzed. Features most important for intermediate cell marker analysis were those that measured the variation in nuclear texture measurements across the slide as well as densitometric features. The most discriminatory contextual features were those that measured the variation of the arrangement of cells within a cluster across the slide. Linear discriminants were calculated using the most important features from both types of analysis; they provided smear classification accuracies of 71-86%.