An AI reading cervix photos spotted precancer better than experts

Researchers at the National Cancer Institute trained a deep learning algorithm on digitized film photographs of the cervix taken during a long-running study in Costa Rica, teaching it to flag the precancerous changes that a human reviewer or a standard lab test might miss.

The algorithm was built and tested on images from 9,406 women followed for up to 7 years, among whom 241 later developed confirmed precancer and 38 developed cancer. It identified precancer with an accuracy score of 0.91, well above the 0.69 scored by expert visual review and the 0.71 scored by conventional Pap cytology.

A photograph of the cervix, rather than a lab sample, is cheap enough to collect where no pathology lab is nearby, which is the real promise: screening that does not wait on infrastructure most clinics do not have. The approach has since been tested on digital and smartphone images in other countries, though it still needs validation beyond the original Costa Rica cohort.

A simple photo outperformed the trained eye.