ABSTRACT
An increasing proportion of the recent cytodiagnostic literature has focused on automation of the Pap smear screening process in hopes of finding a feasible system to aid in the reduction of the number of reported false-negative cases. In a sense, these systems can be thought of as computer-driven sensitivity enhancers for better detection of abnormalities in smeared cervicovaginal specimens. The PAPNET system (Neuromedical Systems, Inc., Suffern, NY) relies on a neural network of artificial-intelligence technology to recognize the complex cellular arrays present in Pap smears, and was originally intended to aid in the identification of morphologically abnormal cells of squamous origin. Herein, we present the results of 61 smears containing a mixture of known diagnostically important benign, dysplastic, and malignant glandular cellular abnormalities which were reviewed by the PAPNET technology. The PAPNET system detected the diagnostic glandular material in 44 of the 45 benign cases reviewed (98% detection rate). In addition, the PAPNET technology identified abnormal cellular material in 15 of the 16 studied smears from patients with malignant/dysplastic morphology (94% detection rate). These data indicate that the PAPNET neural networks are capable of detecting cells with aberrant glandular cytomorphology. In both cases missed by the PAPNET system, the number of abnormal cells per slide was very low, indicating that as with human screeners, the capabilities of this semiautomated method may be exceeded when an extreme paucity of diagnostic cellular material is present in a given slide. Further and larger reviews of glandular abnormalities by automated technologies are needed to assess these systems for their true efficacy at diminishing false-negative cases.
