Background: New technology for computer-assisted screening of cervical smears that uses neural networks could potentially decrease numbers of screening errors and improve productivity. We assessed an interactive automated system (PAPNET) for primary screening of cervical smears.
Methods: In January, 1997, the National Health Service research and development programme sponsored a multicentre trial to investigate the use of PAPNET for classification of routine cervical smears as negative or needing further microscopic review. compared with conventional primary screening. The study complied with international standards for assessment of automated cervical screening systems. 21,700 smears were analysed by the two methods and were classified as inadequate, negative, mild, moderate, or severe dyskaryosis, invasion, glandular neoplasia, and borderline nuclear changes. 2906 abnormal smears and 298 negative smears were sent for independent cytological review (gold standard). We calculated sensitivity and specificity relative to the findings of the independent review.
Findings: Agreement of 89.8% between the two methods was shown for all classifications of smears that were adequate for reporting. The sensitivity was similar for correctly identified abnormal smears on PAPNET-assisted (82%) and conventional screening (83%). PAPNET-assisted screening showed significantly better specificity (77%) than conventional screening (42%) for identification of negative smears. The total mean time for screening and reporting for conventional screening was 10.4 min per smear, and for PAPNET-assisted screening was 3.9 min.
Interpretation: Use of PAPNET-assisted screening could increase quality and productivity. We recommend carefully organised and controlled development projects for the introduction of PAPNET-assisted primary screening of cervical smears.