American researchers have developed an algorithm that can accurately detect precancerous cells in the cervix. In the long term, healthcare staff might be able to screen for this type of cancer, which is particularly prevalent in poor countries, without any training.

Cervical cancer is the fourth most common cancer worldwide, with 570,000 cases per year reported in 2018, says the World Health Organisation (WHO). To improve diagnosis, research could be directed towards artificial intelligence.

Indeed, an algorithm could prove much more accurate than humans and traditional examinations in detecting precancerous cells in the cervix, says a study led by American researchers that was published on Thursday 10^th January in the National Cancer Institute journal.

A disease of poverty

In wealthy countries, the advances made in cervical cancer screening and the vaccination campaigns against papillomavirus (linked to cancer) have reduced mortality rates, but these advances often do not reach poorer countries, which suffered 90% of cervical cancer-related deaths in 2012, says the WHO. "Cervical cancer is now a disease of poverty, of low resources," said Mark Schiffman, a doctor at the National Cancer Institute near Washington, who has been treating this type of cancer for 35 years. "We are trying to find ways that are extremely cheap, extremely easy but very accurate, so that we can attack cervical cancer by vaccine and also a bit later through a simple technique that is cell-phone based or something like it," he continued.

In this way, the doctor and his team have developed an algorithm using an archive of 60,000 images of cervixes taken in Costa Rica in the 1990s. In total, 9,400 women took part in the study and were monitored for several years, 18 years in some cases.

An algorithm that is much more accurate than humans

Once it had been trained on this database of images, the algorithm was able to visually detect precancerous cells with 91% accuracy. A human expert achieved a 69% success rate, and conventional examinations such as cervical smears were 71% successful. "It performed much better than humans looking at those same pictures. It certainly performed a lot better than me," said Mark Schiffman.

In the long term, the aim is to develop this technology over the next three to five years and carry out more clinical trials. Healthcare staff will then be able to screen patients without undergoing any training, simply by taking a photo that will then be analysed by the algorithm.  "I think now we have a possible tool that can go anywhere and not sacrifice scientific quality but actually offer a medically valid screen," concluded the doctor, hopefully.

Allianz Partners