Cervical screening by neural networks

In this issue of Cancer Cytopathology, Kok et al. report their results evaluating cervical smears by neural network screening, and attempted to define more precisely those women whose smears show Grade 1–2 cervical intraepithelial neoplasia (CIN). By defining this borderline category between strictly low-grade and high-grade CIN, the authors endeavored to demonstrate an improved cytohistologic correlation. Because many of these lesions will regress without treatment, the ultimate goal of the authors is to reduce the need for medical intervention. Between 1992 and 1995, the study laboratory screened 133,196 cervical cytologic samples. A total of 83,404 women were screened using the neural network technology, which detected 1128 cases that were interpreted as Grade 1–2 CIN. The number of smears demonstrating Grade 1–2 CIN among 49,792 women who underwent conventional screening was 1108. Thus, the prevalence rate for smears showing Grade 1–2 CIN was nearly double for conventionally screened smears (2.04% vs. 1.15% for smears detected by neural network screening). In the study by Kok et al., the first tissue studies obtained after the abnormal smear demonstrated a concordance with histologic follow-up that was nearly twice that of conventionally screened smears (53.9% vs. 29.2%). While both groups of cases had overinterpretations and underinterpretations of the smears when compared with follow-up tissue studies, overdiagnosis in the cases screened by neural network technology was much lower than that for conventional screening (39.4% vs. 62.4%). Although it may surprise the readers of the journal that an article on screening by neural networks technology was included, the technology remains viable outside the U.S. It is well known that Neuromedical Systems, Inc. (Suffern, NY), the company that developed the neural network imaging system, is under bankruptcy protection by the courts. The intellectual property of the company has been purchased by TriPath Imaging (Burlington, NC) and may be incorporated either whole or in part into future generations of their imaging systems for screening cervical smears. Data from the article by Kok et al. also emphasizes the original thrust of the neural network imaging system, which, in the case of cervical smears, is the ability to detect the rare and significantly abnormal cell as well as the small undifferentiated reserve or parabasal type cell that, when present in small numbers, is very difficult to find conventional screening. This particular focus of neural network technology has been demonstrated in a number of published reports, 171 CANCER CYTOPATHOLOGY