Kidney cancer is one of the most over-treated cancers in Denmark. The available scan images are often unreliable, with one in five CT scans yielding false positives. This means that up to 27 percent of kidney tumor patients undergo painful biopsies and surgeries without having cancer.
To address this, a newly developed AI model is currently being tested at Zealand University Hospital. It surpassed experienced doctors in diagnosing kidney cancer based on scan images. The problem is, however, is that doctors cannot explain the model’s conclusions. This hinders the AI model’s widespread adoption.
In the research and innovation project EXPLAIN ME, funded by the Digital Research Centre Denmark, a team of researchers from the University of Copenhagen, Roskilde University, and the Urology Department at Zealand University Hospital are working to interpret the model’s conclusions.
“Although it is tempting, we cannot simple leave such significant decisions to AI. We need to fully understand its neural patterns from the outset before we can implement it in practice,” says Nessn Azawi, Chief Physician at Zealand University Hospital’s Urology Department and Associate Professor at the University of Copenhagen.
Significant savings for society
As part of the EXPLAIN-ME project, Nessn Azawi and his research team have been working since 2022 to develop explainable artificial intelligence (XAI) that can guide nephrologists on when surgery is necessary, and crucially, explain why.
The 1,000 Danish patients diagnosed with kidney cancer each year rarely show symptoms until the cancer is advanced. The significant diagnostic uncertainty leads to many patients being over-treated. According to Nessn Azawi, AI-based diagnosis could reduce the treatment process by 2-4 weeks and save the healthcare system approximately 15-25 million kroner annually. These positive outcomes would be maximized if the technology is adopted throughout the Nordic region.
“We over-treat around 30,000 kidney cancer patients in the Scandinavian countries. Improving diagnosis would have significant positive effects for both society and the patients,” says Nessn Azawi.
A multidisciplinary effort
Researchers have already tested the AI model at Roskilde University with promising results. The next milestone is to develop a model with a more detailed dataset that can provide nephrologists with accurate kidney cancer diagnoses supported by solid evidence. This has been the focus of PhD student Daniel van Dijk Jacobsen from Roskilde University’s Department of People and Technology for the past two years.
“The challenge is that we don’t know what the model is analyzing when it makes the diagnosis. It’s about identifying the patterns the model detects at the pixel level and then conveying that information to the doctors,” he says.
Thus, it has been essential to work across disciplines, incorporating ethnographic observation studies during patient interactions, participatory design, and ongoing discussions with the medical staff at Zealand University Hospital.
“I find that doctors are enthusiastic about exploring technological possibilities, as they are eager for assistance in achieving more precise diagnostics. They want to be able to compare the patient’s history with the machine’s diagnosis and make decisions based on a better foundation than they currently have,” says Daniel van Dijk Jacobsen.
