Deep-Learning–Based Evaluation of Dual Stain Cytology for Cervical Cancer Screening: A New Paradigm
Annekathryn Goodman · JNCI Journal of the National Cancer Institute · 2020
Cervical cancer is diagnosed in more than 13 000 women, of whom 30% will die in the United States each year (1). Worldwide, more than 3% of both the global burden of cancer incidence and death is attributable to this preventable disease with an annual estimated 570 000 cases and 311 400 deaths (2). Prevention of cervical cancer can be primary through human papillomavirus (HPV) vaccination or secondary through repetitive screening during a woman’s reproductive years. The challenge of screening has been to capture true positives and reduce the false positives that lead to over-testing and increased medical costs. The current recommendation of combined cytology with HPV testing of high-risk subtypes at 5-year intervals has improved the positive predictive value of screening but still leads to over-referrals for colposcopy because of the high prevalence of transient HPV infections (3). An additional challenge for resource-limited regions is the need for a sophisticated infrastructure requiring materials for cytology and HPV testing; processing instruments; highly trained cytopathologists; systems for reporting, tracking, and follow-up; and the lack of health-care insurance in many regions requiring out-of-pocket pay by patients. Wentzensen et al. report on a possible solution to some of the challenges of both screening specificity and infrastructure for clinical delivery of care (4). They report on a cloud-based whole-slide imaging platform of slides stained for both p16 and Ki-67, 2 markers that are closely linked to cervical carcinogenesis. Using a deep-learning classifier for 2 different liquid-based cytology techniques, ThinPrep and SurePath, they noted equal sensitivity and higher specificity compared with both cytology and manual evaluation of dual-stained slides and a reduction of referrals to colposcopy from 60% to 41.9%. Machine learning trains and programs computers to learn patterns from data and is based on sets of mathematical rules and statistical assumptions. The learned program or classifier, an algorithm that sorts data into categories, is then used to predict or identify a pattern. Examples of machine learning in medical research include identification of gene sequences and quantifying molecular variables associated with disease (5). While there has been a fertile evaluation of computer-assisted programs for screening of cytology (6) and a recent report of deep learning for nasal cytology (7), Wentzensen et al. report on the first, to my knowledge, comprehensive analysis of machine learning for cervical cancer screening. Their rationale of developing a deep-learning algorithm for a dual stain of p16 and Ki-67 instead of cytological image identification comes from the observation that the increase in the number of dual-stained cells correlates with increasing severity of histopathology. The authors validated their machine-learning program using 3 large study populations: the Biopsy Study of University of Oklahoma, the Anal Cancer Screening study of HIV positive men, and HPV-positive women from a large screening clinic through Kaiser Permanent Northern California. A limitation of this study is that the test populations had the best possible technical cytology acquisition. With that, they required 2 separate classifiers for ThinPrep and SurePath. One worries that in a less controlled environment of suboptimal examinations and limited supplies that the results may be less robust. It would be important to do a follow-up study with less generalizable populations. The advantage of a deep-learning evaluation of dual-stained slides includes both a reduction on the reliance of trained cytopathologists and the need for proximity of these resources. With a cloud-based location, an analysis could be performed on samples taken anywhere in the world. An important companion system to be developed is a mechanism to deliver images for analysis. Although the authors discuss a courier-based delivery system, this may be impractical in many parts of the world. Mobile phone technology has been used to send images of cervical lesions (8). Consideration should be given to developing technology for sending cytological images by cell phone as well. Conflict of interest: The author has no conflicts of interest to disclose.