Artificial intelligence could detect the start of cervical cancer in women better than trained doctors, a new study claims.
Researchers at the National Cancer Institute (NCI) and Global Good developed a computer algorithm that can identify precancerous changes in a woman’s cervix up to 1.3 times better than standard tests.
Previous tests involved swabbing the cervix with a vinegar solution, but the solution wasn’t able to distinguish between inflammation, infection and precancer.
The team says the AI model is faster, cheaper and more accurate than current measures, and could expand to countries that have limited access to advanced screening methods.
A new study says a computer algorithm can identify precancerous changes in a woman’s cervix up to 1.3 times better than standard tests (file image)
Cervical cancer occurs in the cells of the cervix, the lowermost part of the uterus.
The American Cancer Society estimates that more than 13,000 US women will develop cervical cancer in 2019, which will lead to more than 4,200 deaths.
It is most often diagnosed between ages 35 and 44, but a recent study from the Mayo Clinic found that less than two-thirds of women ages 30 to 65 were not up date with the cervical cancer screenings.
These screenings detect precancerous changes in the cervix before they develop into disease.
Such changes include normal glandular cells transforming into squamous cells, which are more sensitive to the effect of the human papillomavirus (HPV), according to the Canadian Cancer Society.
Another cancer precursor is called adenocarcinoma in situ, which is when abnormal cells are found in the glandular tissue that lines the uterus and cervix.
The current standard for screenings, particularly in areas with few resources, is called visual inspection with acetic acid.
‘The cervix is swabbed with a weakened vinegar solution,’ Jennifer Loukissas, chief of communications in the Division of Cancer Epidemiology and Genetics at NCI, told DailyMail.com.
‘The solution turns HPV white and any cellular changes white while pink is normal. But white can be infection, inflammation or precancer. That’s why it’s so difficult to interpret with the human eye.’
A 2003 study found the test was about 80 percent accurate in correctly identifying those with the disease as positive.
To train the algorithm, the team recruited more than 9,400 women from Costa Rica between ages 18 and 94 years.
More than 60,000 images were taken during cervical cancer screenings between 1993 and 2000 and the woman were monitored over an 18-year period.
The team digitized the photographs and then trained the algorithm to recognize which cervical changes needed treatment and which didn’t need it for the new method called automated visual evaluation.
‘First you get [the algorithm] to recognize the cervix,’ senior author Dr Mark Schiffman of NCI’s Division of Cancer Epidemiology and Genetics, told DailyMail.com.
‘You take 80 percent of the bad and the good and train it to identify “this” is bad and “this” isn’t bad. Then you take an independent 20 percent not related and see if it can tell the difference.’
The team used a statistical model to quantify its results called area under the curve. An AUC of 0.5 means it’s about the same as guessing and an AUC of 1.0 means perfect accuracy.
The AI tool, when identifying precancer, achieved an AUC of 0.91, about 1.3 times better than the review of human experts, which achieved an AUC of 0.69.
Pap tests also performed worse, achieving an AUC of 0.71.
‘With a human expert, it’s open to interpretation and, even though the brain is good at detecting patterns, the neural networks are better,’ said Loukissas.
‘It would make screening and triage more cost-effective in places where they currently aren’t.’
For the future, the researchers hope to put the program onto a cellphone so that any woman in any country would be able to take a photo of their cervix and get their results back quickly.
‘Cervical cancer has become a health disparity issue,’ said Dr Schiffman.
‘In Malawi, five percent of women get cervical cancer in their lifetime. In the US, it’s one-tenth of that.
‘So we have something very simple, very inexpensive that can go anywhere where the HPV vaccine is not reaching people and where good testing isn’t reaching people.’
The study comes on the heels of a growing body of research that has found that artificial intelligence can improve upon the work of physicians.
This week, two separate studies from the Mayo Clinic and Stanford University found that AI was able to identify irregular heart rhythms and the early stages of heart failure.
In some cases, the AI model did so up to 15 percent more accurately than EKG results analyzed by board-certified cardiologists.