Skip to main content
Premium Trial:

Request an Annual Quote

Karolinska-Led Team Develops AI Tool to Diagnose, Grade Prostate Cancer Cases

Premium

NEW YORK – Investigators at the Karolinska Institutet in Sweden have developed a new artificial intelligence tool that they claim can evaluate prostate cancer biopsies at a level on par with pathologists.

The group aims to validate the tool, called OncoWatch, in a study involving nine countries this year. It also hopes to achieve a CE-IVD mark for the test in 2020. A paper detailing the development of OncoWatch was published last week in The Lancet Oncology.

According to Martin Steinberg, the project leader for OncoWatch, the group last year received €2.7 million ($3 million) to validate the tool in an international trial. EIT Health, the European innovation network, awarded the funding. The EIT Health-funded project, which commenced in January 2019, is slated to run through 2021. The first year of the project was spent preparing the study, Steinberg said.

The main concept behind OncoWatch is to alleviate the burden on pathologists in assessing prostate cancer biopsies. "Pathologists are becoming a rare species," noted Steinberg. "There are fewer of them and they have a greater workload."

There are other pressures on clinicians in regard to prostate cancer cases. Steinberg noted a lack of specificity in prostate-specific antigen testing, a risk of infection connected with invasive biopsies and surgery, as well as multiple new therapies coming on the market coupled with a lack of markers to determine which patients would most likely benefit from them.

Another issue is subjectivity. Results often vary from lab to lab, and between pathologists, and patients are often over- or undertreated because of discordant pathology results. For all of these reasons, the Karolinska researchers turned to AI to develop a new solution.

Much of the work has been grounded in the development of Stockholm3, a blood-based test that predicts the risk of aggressive prostate cancer by analyzing five protein markers, more than a hundred genetic markers, and clinical data. The Karolinska researchers introduced Stockholm3 as a laboratory-developed test years ago, and the test is now in use in Sweden, Norway, Denmark and Finland, with about 20,000 tests sold last year, Steinberg said. He added that Synlab International and Unilabs, two of the largest laboratory chains in Europe, will start offering Stockholm3 this quarter.

Using the data set from the development of the Stockholm3 test, the Karolinska researchers set about creating an AI tool for prostate cancer that could lessen the burden on pathologists. "In the end, what we had from Stockholm3 was that 60,000 men participated in the study, and we had 7,000 biopsies of men evaluated by the same pathologist," Steinberg said. "Without the Stockholm3 trial, we wouldn't have this unique data set."

Martin Eklund, an associate professor of biostatistics at Karolinska, agreed. "All of those biopsies were available, and we could scan them using digital pathology scanners," Eklund said. "Using that data we could train AI for cancer detection and how to assess how much cancer is in biopsy," he said.

The most difficult task for the developers of OncoWatch has been training the AI tool to assign a Gleason score to an individual biopsy. The lower the score, the more likely a cancer will be slow-growing. The higher the score, the more immediate need for surgery or therapy. "The consequence of the Gleason score for the future management of the man is fundamental," Eklund said. "It was very important to get that right."

The Lancet Oncology paper further details the development of the test. While the study involved investigators from around the globe, Eklund credited partners at the University of Tampere in Finland in particular with helping to create the tool. Kimmo Kartasalo, a PhD student at Karolinska whose primarily affiliation is at Tampere, was "instrumental to the project," said Eklund.

In the study, the investigators digitized 6,682 slides from needle core biopsies from 976 randomly selected participants aged between 50 and 69 sourced from the Stockholm3 diagnostic study. An additional 271 biopsies obtained from 93 men were folded into the study.

These were used to train AI to predict the presence, extent, and Gleason grade for an independent data set of 1,631 biopsies from 246 men from the Stockholm3 study and an external validations set of 330 biopsies from 73 other men. The results were highly accurate — about 95 percent concordance in determining benign from malignant tissue and the extent of the cancer. The assignment of Gleason grade was also concordant with pathologists.

Based on these findings, the authors believe the system could act as a safety net, reducing the assessment of benign biopsies, automating the determination of cancer extent, and provide scoring comparable to pathologists. "An AI system with expert-level grading performance might contribute a second opinion, aid in standardizing grading, and provide pathology expertise in parts of the world where it does not exist," the authors wrote.

Yet while they see a clinical path for OncoWatch, there is much to be done before it is adopted widely.

Validating the test

According to Steinberg, the EIT Health-backed OncoWatch Project will soon commence a study of the test across nine countries, which is slated to wrap up by the end of 2020. As part of the study, participating centers will train the AI tool to recognize cancer in biopsies taken from various labs, with different kinds of digital pathology scanners, and with rarer growth patterns.

Participating centers include Karolinska University Hospital and Capio St. Göran Hospital in Sweden; the Medical University of Lodz in Poland; Taunton and Somerset NHS Foundation Trust in the UK; University Hospital Erlangen in Germany; Stavanger University Hospital in Norway; Aarhus University Hospital in Denmark; Mehiläinen, a private healthcare provider in Finland; and Synlab locations in Italy and Switzerland.

The investigators have asked for samples from 200 men from each site, which translates to about 2,000 needle biopsies from each participant, or 20,000 needle biopsies in total.

"What we have shown so far is that it's possible to create an AI that can do diagnostics and Gleason grading on the level of expert pathologists," said Eklund. "To turn it into a product, more work is needed," he said, including validating the tool across various scanners at different labs on samples with unusual morphologies.

"We have used different digital pathology scanners in this work, and we know depending on what scanner we use, the predictions can change," he said. One solution might be to recommend the tool be used with a particular scanner. For its preliminary work on OncoWatch, the researchers have used Philips IntelliSite Pathology Solution Ultra Fast Scanner.

Steinberg noted that the Philips scanner has already attained a CE-IVD mark and that labs chains like Synlab and Unilabs are investing in the scanner. "It's really timely," he said of the work. He noted that Philips is collaborating with the researchers to streamline the use of OncoWatch with its scanner, though the relationship is not exclusive.

Another option is to tweak the tool so that it can be used across most scanners. "We know it's possible, but we have to show that, as well," said Eklund. "That is one thing we are working on heavily within the project funded by EIT Health." Complicating the process are variables introduced by the labs themselves, which process specimens in different ways.

"Staining is different, cutting is different, and so we need to investigate [the use of the tool] across a wide range of different labs to provide high accuracy;" Eklund said.

As part of the path to market, the investigators will also obtain a CE-IVD mark for OncoWatch. Steinberg said that OncoAlgorithms, an Uppsala, Sweden-based company, will undertake the CE-IVD marking process under the current In Vitro Diagnostic Directive by the end of 2020. Europe is currently transitioning to a new IVD Regulation, which will come into force in 2022. By 2022, Steinberg said he expects the test will be cleared under the new IVDR.

Steinberg is the chairman of privately held OncoAlgorithms, which currently employs five people. The same company is looking to obtain a CE-IVD mark for the Stockholm3 test in the next year. Steinberg said that a number of labs in North America are working to introduce the Stockholm3 test in 2020 — one in Indiana, another in British Columbia. There has also been some interest in Stockholm3 from China. OncoWatch might also have an international perspective. Once the company gains a CE-IVD mark for the AI test it will start promoting it more broadly.

"Based on user interviews we believe that the AI product would have an excellent fit to the US market," Steinberg said.