Estonian AI helps radiologists detect missed kidney cancers

Tartu University Hospital has successfully tested artificial intelligence developed by Estonian startup Better Medicine, detecting kidney cancer in four patients whose tumours had gone unnoticed by experienced radiologists, ERR reported.

“We have created artificial intelligence that is now like a second pair of eyes for doctors, looking for kidney cancer,” Better Medicine chief executive Priit Salumaa said on the Vikerhommik programme.

The company’s algorithm scans large sets of medical images for changes that are too small to detect with the naked eye or are hidden behind other tissue. When doctors and the system worked together, cancer detection sensitivity reached 99.2 percent, meaning they identified almost all lesions between them.

However, the system is not infallible. Like people, artificial intelligence can make mistakes and may raise a false alarm in places where there is no cancer, Salumaa said.

The hidden nature of kidney cancer makes the technology useful. The disease usually causes no symptoms in its early stages. Because the kidneys are located deep in the abdominal cavity, a small tumour does not generally cause pain or other noticeable complaints. As a result, countries do not have dedicated screening programmes to systematically monitor people in risk groups.

Doctors most often discover kidney tumours by chance. A patient may arrive at an emergency department, for example, with a broken bone or unexplained abdominal pain. Doctors may then perform a computed tomography scan to rule out internal injuries. The scan produces detailed cross-sectional images, and the AI analyses these routine image series in the background, looking for signs of other problems.

The Estonian algorithm also stands out internationally because, so far, no other artificial intelligence medical device that searches for kidney cancer in abdominal CT scans has received medical device certification and CE marking. Because the programme meets the European Union’s strict safety requirements, doctors may officially use it to support their decisions.

Salumaa said he had initially feared doctors would react negatively because the machine could expose human inattention. Specialists, however, welcomed the solution. When the four cases were identified, they were immediately acted on and notifications were sent to the treating doctors, he said.

Five years of continuous work preceded the solution’s launch. Salumaa said the most difficult challenge was gaining access to the data needed to develop the artificial intelligence, as well as navigating the regulatory framework.