Abstract B58: Assessment of models for combining colposcopy features for cervical cancer diagnosis

Abstract

Abstract B58 Despite the introduction of vaccines to prevent two types of high-risk oncogenic strains of the human papillomavirus (HPV), screening and diagnosis will still be important for the prevention of cervical cancer. Colposcopy is used to visualize the cervix to identify abnormal lesions for biopsy, which in turn determines disease severity and management. Recent studies have shown that colposcopy is not as sensitive as once believed, and colposcopy is a critical step in the prevention of cervical cancer. The HPV vaccine will likely result in a decline in the number of cases referred to colposcopy due to HPV 16 and 18. HPV 16 or 18 infection confers a high risk of precancer or cancer, and may represent cases that are easier to diagnosis. A reduction in the number of these cases may result in colposcopy becoming more challenging, further highlighting the need to optimize colposcopy. During colposcopic interpretation, physicians use the features of color, margins, mosacisim, and punctation, to determine a diagnosis for a patient and which areas to biopsy. How the features are combined to make these decisions significantly impacts the performance of colposcopy and biopsy as a diagnostic tool for cervical cancer. With the changing population referred to colposcopy, it becomes even more important to understand how to best utilize colposcopic features and colposcopy as a diagnostic tool. While we cannot predict the population that will be referred to colposcopy after widespread introduction of the HPV vaccine, patients diagnosed with atypical squamous cells of undetermined significance (ASCUS) and low-grade squamous intraepithelial (LSIL) Pap smears represent challenging populations for colposcopic diagnosis. We analyzed the performance of colposcopy using image annotations from 21 colposcopists evaluating digitized Cervigrams, which are magnified images of the cervix, from a sub-study of the ASCUS LSIL Triage Study (ALTS). We randomly selected one-third of the patients enrolled in the sub-study, stratified by final pathology and HPV status, for this evaluation. Computerized models for combining colposcopic features may improve the performance of colposcopists. We explore the performance of different models for combining colposcopic features, including a summed index score similar to the Reid index and logistic regression models, in comparison to the gold standard of diagnosis from pathology, to determine if better methods are available for combining colposcopic features than the colposcopic diagnosis provided by the colposcopist. The colposcopic diagnosis is the implicit aggregation of features performed by the colposcopist. To assess the performance of the various models, we use prediction probability as the figure of merit. Prediction probability is a generalization of the area under the receiver operating characteristic curve from binary to multi-level disease outcomes, allowing us to assess outcomes of cervical intraepithelial neoplasia 0, 1, and 2+. There is variability in performance among colposcopists, with a handful performing better than either the summed index score or logistic regression models. The driving forces of the variability among colposcopists need to be further elucidated to determine if there is additional information that should be used in colposcopy to improve the overall performance of colposcopists in general practice. Citation Information: Cancer Prev Res 2008;1(7 Suppl):B58.