Automated evaluation of digital images of the cervix may have potential in point-of-care cancer screening, according to results of a proof-of-concept study supported by the National Institutes of Health. Published in The Journal of National Cancer Institute, the research showed that a computer could be taught to distinguish cervical precancer/cancer from normal cells in images using a “deep learning” visual evaluation algorithm.
The images in the study were cervigrams, pairs of cervical photographs that are visually screened for cervical cancer and precursors. The method, called cervicography, has been discontinued but was in use in Guanacaste, Costa Rica during the period of study—1993 to 2000. The population longitudinal cohort for this research was 9406 women aged 18 to 94 in that geographic area who were followed for 7 years.
The participants were screened with cervicography and other methods and their precancers were histologically confirmed. Tumor registry linkage identified cancers up to 18 years. Images from the patients’ cervigrams were used to train an automated visual evaluation algorithm. The authors designed the algorithm to detect the cervix within an image and predict the probability that the image represented cervical intraepithelial neoplasia (CIN) 2+. The resulting image prediction score (0-1) could be categorized to balance sensitivity and specificity for detection of precancer/cancer.
Testing of automated evaluation of the cervigrams showed that the system it was more accurate in identifying precancer/cancer cases than the original cervigram interpretations (AUC 0.91, 95% confidence interval [CI] 0.89 ot 0.93 vs AUC 0.69, 95% CI 0.63 to 0.74; P < .001). The automated technique also was more accurate than conventional cytology (AUC 0.71; 95% CI 0.65 to 0.77; P < .001). A single visual screening round restricted to women aged 25 to 49 could identify 127 of 228 precancers (55.7%) diagnosed cumulatively in the adult population while referring 11% for management.
The authors said their results “support consideration of automated visual evaluation of cervical images from contemporary digital cancers,” to possibly expand point-of-care cervical screening. They believe it also has potential as a triage method for women who test human papillomavirus (HPV)-positive, if HPV is the primary screen and could reduce the need for speculum examinations. They noted, however, that this study had a small number of cases from a single cohort and that the algorithm ideally should be trained on more definitive precancer cases, such as CIN3 and adenocarcinoma in situ.
As an extension of this research, the authors are transferring automated visual evaluation to images from contemporary phone cameras and other digital image capture devices to create an accurate and affordable point-of-care screening method that would support a recently announced World Health Organization initiative to accelerate cervical cancer control.