The PAPNET system is an automated interactive instrument for analysis of conventional (Papanicolaou) cervical smears. The instrument, described in this paper, introduces several important innovations to cytology automation. The cell selection system is composed of two stages: an algorithmic classifier, followed by a trained neural network allowing for great flexibility and precision in recognition of abnormal cell images. Contrary to other attempts at cytology automation, this machine does not attempt to diagnose cell abnormalities. Instead, it is interactive, leaving the assessment of the cells displayed on a high-resolution video screen to trained human observers. The slides judged to contain abnormal cells or to be inadequate are referred for a second microscopic review. Two versions of the instrument (Alpha and Beta) were evaluated in several modes. Initial testing was performed on archival smears with known, histologically confirmed neoplastic lesions of the uterine cervix. These lesions comprised the entire spectrum of abnormalities, from low-grade lesions to invasive cancers of several types. The Alpha machine displayed recognizable abnormal cells in 97% of the 201 cases, and the Beta machine displayed such cells in 97.2% of 176 cases. The Beta instrument was subsequently tested on 500 sequential archival cervical smears that had been previously subjected to a rigorous quality control. One hundred forty smears (28%), which either displayed atypical cells or were considered "inadequate," were referred for further rescreening. Fifteen of 16 previously diagnosed neoplastic smears were appropriately identified with the help of the machine. The one missed case contained a single cluster of vacuolated cancer cells from an endometrial carcinoma. As a result of PAPNET-triggered review, three new cases of low-grade squamous intraepithelial lesions view, three new cases of low-grade squamous intraepithelial lesions (SIL) came to light in previously negative smears; three additional cases, previously classified as atypical, were also reclassified as SIL, for a net gain of six neoplastic abnormalities. In two additional atypical cases, colposcopic follow-up was recommended, even though the diagnosis was not modified. Two cases of cervical intraepithelial neoplasia, represented by tiny single clusters of abnormal cells missed on original screening, quality control, and on machine rescreening, came to light on second review of the residual 360 cases. The initial experience with the PAPNET system suggests that the instrument may be valuable in quality control and may assist in significantly reducing false-negative cervical smears in an efficient and timely manner. Further testing of the instrument on a much larger number of cervical smears is in progress.

No abstract

No abstract