NEURAL NETWORK-ASSISTED ANALYSIS AND MICROSCOPIC RESCREENING IN PRESUMED NEGATIVE CERVICAL CYTOLOGIC SMEARS - A COMPARISON

OBJECTIVE: To compare cytologists' defection of abnormalities when usi ng neural network-assisted (NNA) review, as employed by the PAPNET Tes ting System and to compare the effectiveness of this mode of review to that of unassisted, conventional rescreening of cervical smears initi ally diagnosed as negative. STUDY DESIGN: The study was undertaken as part of a multicenter clinical trial involving over 10,000 smears from 10 investigation sites (9 academic institutions and 1 private laborat ory). Using a subset of ''negative'' control smears from three univers ity laboratories, the false negative detection yields of NNA review (p erformed using the PAPNET System) and conventional microscopic rescree ning (performed as part of routine quality control practice) were comp ared. The false negative detection yield was defined Its the percentag e of rescreened negatives reclassified as abnormal. RESULTS: The resul ts demonstrate that using NNA review, the detection yield of false neg ative smears, as a proportion of negative smears reexamined, is statis tically significantly greater than that obtained using conventional qu ality control rescreening. The false negative yield generated using NN A analysis was 6.2% (142/2293) versus 0.6% (82/13761) for conventional rescreening. A statistically significant improvement in identificatio n of abnormality is observed for NNA review as opposed to unassisted r escreening despite constraining the comparison in the following ways: (1) comparing the yields of rescreening of negative smears obtained fr om the same time intervals for both methods, (2) comparing the yields of rescreening of negative smears obtained from the years after the Cl inical Laboratory Improvement Act (1990 and 1991) for both methods, an d (3) disregarding the identification of atypical squamous cells of un determined significance/atypical glandular cells of undetermined signi ficance cases and comparing only the identification of squamous intrae pithelial lesions using the two methods. CONCLUSION: Using neural netw ork-assisted review, cytologists uncovered a significantly higher prop ortion of previously undetected cervical abnormalities per smear reexa mined than they did using unassisted, conventional rescreening.