Automated evaluation of p16 immunohistochemistry for diagnosis of cervical precancer
MetadataShow full item record
Purpose: Cervical cells transformed by high-risk strains of Human Papilloma Virus (HPV) overexpress the p16 protein, a cyclin-dependent kinase inhibitor that indicates activation of the E6 and E7 oncogenes. The Lower Anogenital Squamous Terminology (LAST) Standardization Project for HPV-Associated Lesions recommends the adjunctive use of p16 immunohistochemistry (IHC) in cervical biopsies to support diagnosis of Cervical Intraepithelial Neoplasia (CIN), which in some cases progresses to cervical cancer. However, evaluation of p16 expression is subjective. Development of automation that can quantify p16 expression in biopsies offers efficiency, accessibility, and objectivity in guiding clinicians in their management of women with suspected cancer precursors. We sought to evaluate the performance of artificial intelligence in assessing biopsies based on level of p16 expression. Methods: 251 biopsy specimens were collected from women with abnormal cervical cytology screening. These biopsies underwent p16 IHC and evaluation by a pathologist and were used to train and validate an initial deep-learning algorithm. After epithelial segmentation and p16 quantification, different thresholds of p16 expression were evaluated and correlated with disease status. Results: A threshold of 15 or more segments that demonstrate p16-staining in 70% or greater of the vertical extension of the epithelium produced a sensitivity of 82.26% and a specificity of 86.00% in identifying CIN2 or more pathogenic lesions. Conclusions: Current efforts are being made to further refine the algorithm and select the optimal threshold of p16 expression that correctly identifies CIN2+ biopsies. Artificial intelligence provides a reliable and promising avenue in the assessment of cervical biopsies, especially in low-resource settings where there is not ready access to pathologists and cervical cancer is a more prevalent phenomenon.