New tech's potential to identify high-grade cervical dysplasia may be a boon to low-resource settings
Hu L, Bell D, Antani S, et al. An observational study of deep learning and automated evaluation of cervical images for cancer screening. J Natl Cancer Inst. 2019;doi:10.1093/jnci/djy225.
When cervical screening tests like cytology and HPV testing show abnormal results, colposcopy often is recommended. The goal of colposcopy is to identify the areas that might harbor a high-grade precancerous lesion or worse. The gold standard in this case, however, is histology, not colposcopic impression, as many studies have shown that colposcopy without biopsies is limited and that performance is improved with more biopsies.3,4
Visual inspection with acetic acid (VIA) is an approach used often in low-resource settings where visual impression is the gold standard. However, as with colposcopy, a visual evaluation without histology does not perform well, and often women are overtreated. Many attempts have been made with new technologies to overcome the limitations of time, cost, and workforce required for cytology and histology services. New disruptive technologies may be able to surmount human limitations and improve on not only VIA but also the need for histology.
Novel technology uses images to develop algorithm with predictive ability
In a recent observational study, Hu and colleagues used images that were collected during a large population study in Guanacaste, Costa Rica.5 More than 9,000 women were followed for up to 7 years, and cervical photographs (cervigrams) were obtained. Well-annotated histopathology results were obtained for women with abnormal screening, and 279 women had a high-grade dysplastic lesion or cancer.
Cervigrams from women with high-grade lesions and matched controls were collected, and a deep learning-based algorithm using artificial intelligence technology was developed using 70% of the images. The remaining 30% of images were used as a validation set to test the algorithm's ability to "predict" high-grade dysplasia without knowing the final result.
Findings. Termed automated visual evaluation (AVE), this new technology demonstrated a very accurate ability to identify high-grade dysplasia or worse, with an area under the curve (AUC) of 0.91 from merely a cervicogram (FIGURE). This outperformed conventional Pap smears (AUC, 0.71), liquid-based cytology (AUC, 0.79) and, surprisingly, highly sensitive HPV testing (AUC, 0.82) in women in the prime of their screening ages (>25 years of age).
WHAT THIS EVIDENCE MEANS FOR PRACTICE
Colposcopy remains the gold standard for evaluating abnormal cervical cancer screening tests in the United States. But can we do better for our patients using new technologies like AVE? If validated in large-scale trials, AVE has the potential to revolutionize cervical cancer screening in low-resource settings where follow-up and adequate histology services are limited or nonexistent. Future large studies are necessary to evaluate the role of AVE alone versus in combination with other diagnostic testing (such as HPV testing) to detect cervical lesions globally.
Colposcopy remains the gold standard for evaluating abnormal cervical cancer screening tests in the United States. But can we do better for our patients using new technologies like AVE? If validated in large-scale trials, AVE has the potential to revolutionize cervical cancer screening in low-resource settings where follow-up and adequate histology services are limited or nonexistent. Future large studies are necessary to evaluate the role of AVE alone versus in combination with other diagnostic testing (such as HPV testing) to detect cervical lesions globally.
Continue to: Data offer persuasive evidence...