Enhancing Cervical Cancer Screenings with Artificial Intelligence

The potential for artificial intelligence to advance medical imaging and radiology has been well-documented in healthcare research.

From earlier detection and more accurate assessments, to less expensive testing and improved workflows, AI and machine learning tools have demonstrated their ability to transform the medical imaging field.

In the realm of cancer care, these potential advances are even more pronounced.

The complexity of cancer, as well as the enormous volumes of data providers need to consider when treating cancer, makes this area well-suited to the abilities of AI.

In particular, researchers have increasingly sought to use AI and machine learning tools to improve cervical cancer screening – a process that, despite advancements, is still susceptible to error.

“Over the last decade, there have been major developments in the field of cervical cancer and cervical cancer screening. The recognition that human papillomavirus (HPV) is a primary cause of cervical cancer has led to the development of not only vaccines, but also HPV tests,” Nicolas Wentzensen, MD, PhD of NCI’s Division of Cancer Epidemiology and Genetics, told HealthITAnalytics. 

To determine whether patients may need colposcopy, biopsy, or treatment, providers may also administer Pap cytology tests. These tests require specially trained laboratory professionals to analyze stained slides to look for abnormal cells, allowing providers to identify precancers before they progress to cancer.

“While Pap cytology tests have reduced cervical cancer incidents over the last decade, this approach is very subjective and prone to false-positive findings,” Wentzensen said.

Dual-stain testing has emerged as a way to more accurately predict whether a woman with a positive HPV test has precancerous cervical changes. The test measures the presence of two proteins (p16 and Ki-67) in cervical samples, and studies have shown that this method is more accurate than Pap cytology.

However, the manual dual-stain test has a subjective component, in that an expert has to look at the slide to determine the results. Wentzensen and his team set out to discover whether a fully-automated dual-stain test could match or exceed the performance of the manual approach.

The researchers built a whole-slide deep learning imaging platform that could determine if any cervical cells were stained for both p16 and Ki-67. The group found that the AI-based dual-stain test had a lower rate of positive tests than both Pap cytology and manual dual-stain, with better sensitivity and higher specificity than Pap cytology.

“Our approach allows for the first time to have a fully automated, non-subjective pathway to cervical cancer screening in high-resource settings,” Wentzensen said.

“What we've seen is that the performance of the artificial intelligence approach is substantially better than the current standard. With higher sensitivity and higher specificity for detecting cervical pre-cancers, we can refer fewer women to further diagnostic evaluation. That’s a great trade-off – finding more disease but putting a lesser burden on patients.”

A critical component of this project was gathering the data necessary to train and develop the AI model, Wentzensen noted.

“One important part of developing such algorithms is to use really high-quality datasets for training and then use completely independent studies to validate the algorithms. That's a key requirement of designing these machine learning tools,” he said.

“It is also important to have a large amount of data, because these algorithms typically require large datasets to learn and produce accurate results.”

Researchers also developed the tool to ensure that it can be used by providers in any organization.

“One great feature of our approach is that it is cloud-based, so the algorithm is not confined to a specific location. The samples can be collected and sent to a central laboratory, where the staining and the scanning is conducted, and then the evaluation happens in the cloud and it can be accessed from any place,” Wentzensen said.

“This makes the technology available to users even without having these very high-level laboratories that are needed to conduct the scanning and the evaluation of those large data sets.”

The research team will continue to refine and develop the model to improve cervical cancer screening.

“We are now extending this iteration with even larger data sets, and we will also evaluate it in different populations both across the US and internationally. We will use slides that have already been collected in other studies, but we are also currently conducting new studies where we can evaluate the automated tool stain approach,” Wentzensen explained.

“The other important question we want to answer is, how long does it take for this test to predict the presence or absence of precancerous cells? Because that will tell us how much reassurance a negative test can give. We would like to reassure women immediately if they're at very low risk and identify those who need treatment immediately, and that's our next goal.”

The team also plans to compare the manual and automated approaches with a third intermediate method, where they use the AI algorithm to pre-screen the slide and then present the most concerning cells to a human expert.

“This is what we call an assisted evaluation. It keeps the observer, or whoever is doing the slide assessment in the process, but it uses the power of AI to accelerate the process and make sure that cells are not missed. This could be a transitional phase towards the full implementation of the automated approach,” Wentzensen said.

Looking ahead, AI and machine learning will only become further integrated into clinical practice, with the potential to improve processes and patient care.

“We've seen a number of examples where the AI approach can match the manual evaluation, and our tool is an example of an automated approach that surpasses the manual evaluation in clinical performance. You can already get the benefits of efficiency and reproducibility if you match the manual evaluation, but we show that you can improve the performance even more,” he concluded.

“There are many future applications where this can be important, including diagnostic evaluations, quality control, and process acceleration. There is a wide range of potential applications for AI, and there are already a number of good examples of how the technology could change current clinical practice.”