Fundamental study of automatic cyto-screening for uterine cancer. I. Feature evaluation for the pattern recognition system.

A basic study was carried out to determine the parameters of a pattern recognition system for the automatic assessment of cytologic cell samples. Various cell features were extracted, whose combinations were evaluationed by an "ambiguity function". It was shown that the highest reliability can be obtained with a combination of the features of nuclear staining, nuclear area, area of cytoplasm, nuclear/cytoplasmic ratio, nuclear shape and chromatin pattern. However, recognition of the nuclear edge and chromatin patterns is complicated and makes automation difficult. Even if these two features are omitted, false positives do not exceed 20 per cent. Consequently screening of abnormal cells can be carried out by image recognition procedures by the use of a computer.