Neural network processing of cervical smears can lead to a decrease in diagnostic variability and an increase in screening efficacy: a study of 63 false-negative smears.

A realistic approach for decreasing the number of erroneous diagnoses plaguing cervical cytology screening is to try to reduce the amount of nondiagnostic visual information. The neural network of PAPNET selects 128 cytological views from the routinely prepared smear which in digitized form can be displayed on a high-resolution videoscreen. From these 128 videotiles the abnormal ones can be selected by the diagnostician and brought together on the "summarizing videoscreen" containing 16 tiles. Thus, the diagnostic information can be further condensed. This facilitates the proper interpretation of the diagnostic cell material dispersed over the smear. A series of 63 false-negative smears were rescreened twice conventionally and twice using the PAPNET-assisted method. We found that, using PAPNET, the screening efficacy increased and the diagnostic variability decreased. The PAPNET in particular proved to be superior for smears containing few abnormal cells and cases of malignancies of the reserve cell lineage.