Use of computer-assisted rescreening as an ancillary tool to subclassify AGUS cervical smears.

Abstract

A substantial percentage of women with a diagnosis of atypical glandular cells of undetermined significance (AGUS) on cervical smears harbor a significant squamous or glandular, preneoplastic or neoplastic lesion on subsequent follow-up. Attempts to subclassify AGUS smears by conventional methods have had mixed results. To determine whether subclassification of AGUS cervical smears using computer-assisted rescreening based on the neural network would improve correlation with subsequent histologic follow-up, 91 cervical smears, conventionally diagnosed as AGUS without concomitant squamous lesions, were subjected to analysis by a computer-assisted automated screening system. Computer-generated images were evaluated by a cytotechnologist without the knowledge of the histologic outcomes. Prior to manual review, each case was classified as either within normal limits, no review required; or abnormal, review required. Based on the degree of abnormality, the latter category was further subclassified into either low probability or high probability of abnormality. The results of the computer-assisted reclassification were then compared with the histologic follow-up of all patients. Thirty-three cases (38.8%) had a significant lesion on histologic follow-up. The lesions included 4 CIN I, 7 CIN II/III, 12 endocervical adenocarcinomas (ACA), and 10 endometrial ACA. Based on computer-generated images, 65% of the smears that were triaged as high probability of abnormality, 11.5% that were triaged as low probability of abnormality, and 10.5% that were triaged as within normal limits had a significant lesion on subsequent follow-up. We conclude that computer-assisted rescreening aids in the triage of AGUS smears and that computer-assisted rescreening based on the neural network or other algorithms may be a useful ancillary tool for subclassifying AGUS cervical smears.