AbstractPurpose An emerging application of artificial intelligence (AI) in cervical cytology is its use as an independent screener for triaging cases that do not require human review. The aim of this study was to assess the possibility of clinical practise. Materials and Methods AICyte alone with a 50% negative cut-off is evaluated with detailed analysis of any AICyte negative cases originally diagnosed as atypical squamous cell of undetermined significance or worse (ASC-US+/AI-), including re-examination by three pathologists with and without AICyte assistance and follow-up histologic correlation. Results Of 80899 Pap tests, only 390 (0.48%) were ASC-US+/ AI-. Of these, 379 (97.18%) were originally diagnosed as ASC-US with no cases of high-grade intraepithelial lesion (HSIL) or squamous cell carcinoma. AICyte alone demonstrated 92.11% sensitivity and 98.98% negative predictive value (NPV) for an original ASC‐US+ diagnosis, which increased to 99.44% sensitivity and 99.97% NPV for detecting LSIL+. Upon review of ASC-US+/AI- cases using microscopy by three pathologists, negative for intraepithelial lesion or malignancy (NILM) was interpreted in 46.41%, 40.77%, and 47.95% of cases, and ASC-US in 49.74%, 52.31%, and 46.67% of cases. Review of the cases using AICyte assistance showed NILM reporting rates of 79.74%, 68.46%, and 58.97%. None of the AICyte-negative cases reached consensus among the three pathologists for classification as HSIL, atypical squamous cells - cannot exclude high-grade squamous intraepithelial lesion (ASC-H) using either method, indicating no obvious abnormal Pap cases in the AICyte-negative group, aside from the ASC-US, rare low grade squamous intraepithelial lesion (LSIL) categories. A total of 151 cases (38.72%) were interpreted as NILM by both methods. Of 98 ASC-US+/AI- cases with histologic follow-up, 2 cases showed cervical intraepithelial neoplasia 2 (CIN2) with mixed Pap cytological diagnoses of ASC-US by original diagnosis, NILM by microscopy, and one ASC-US/one NILM using AICyte assistance. One case of endometrial carcinoma was noted in the follow-up, which was classified as atypical glandular cells (AGC) in original, consensus interpretation in both microscopy and AICyte-assistance. For all three pathologists, reading times were significantly lower using AICyte assistance compared to microscopy. Conclusions AICyte can safely function as an independent screener to triage cervical cytology cases not needing further review, reducing the workload by half. It can also function as an assistive tool to improve efficiency and accuracy of cervical cytology interpretation.
Dong et al. (Wed,) studied this question.