Artificial intelligence (AI) can do more than just write lyrics and create songs. For example, AI algorithms can be trained to detect cancer at an early stage. But dare we blindly trust the observations of a computer that a radiologist cannot see with the naked eye?
Het werkt eigenlijk heel simpel: een computer krijgt foto’s van CT-scans te zien. Voor elk plaatje waarop hij bepaalde afwijkingen detecteert, wordt de computer ‘beloond’. Na dit tienduizenden keren te hebben herhaald, heeft de computer geleerd wanneer een mogelijke tumor op de scan te zien is. Hij is zó goed getraind, dat hij er zelfs beter in is dan de mens.
Dit is slechts een van de weinige toepassingen van AI die er op dit moment in de zorg zijn. Maar deze techniek is veelbelovend en kan straks mogelijk veel sneller en nauwkeuriger dan radiologen vaststellen of iemand een kwaadaardige tumor in zijn lichaam heeft. En hoe sneller potentiële kanker wordt ontdekt, hoe beter het vaak te behandelen is.
Deze week werd een nieuwe studie naar zo’n AI-toepassing gepubliceerd in eBioMedicine, de medische tak van het wetenschappelijke tijdschrift The Lancet. De onderzoekers bouwden een AI-algoritme dat in staat is om in een vroeg stadium ‘kankerknobbeltjes’ op longen te detecteren.
“In feite kun je deze techniek voor elke vorm van kanker toepassen”, zegt Mireille Broeders. Zij is als hoogleraar Pesonalized Cancer Screening verbonden aan het Radboudumc. “Dat komt doordat zo’n algoritme informatie uit zo’n beeld weet te halen wat wij als mensen niet kunnen zien.”
Krijg een melding bij nieuwe berichten
AI cancer research took off thanks to self-learning computers
In the study eBioMedicine is one of many studies conducted on the applications of AI in cancer care. According to Jonas Teuwen, the Netherlands is at the forefront of research on these types of algorithms. Teuwen is affiliated with the Netherlands Cancer Institute at the Antoni van Leeuwenhoek Hospital in Amsterdam. “In the Netherlands, we do these kinds of large-scale projects,” he says.
AI and cancer research has been around for much longer, but has only really grown in recent years, Broeders and Teuwen say. Broeders: “As soon as deep learning started to play a role, it really took off.” Deep learning is the technique that allows computers to learn from large amounts of raw data (think of the example of CT scans).
“It’s hard to say what phase we’re in right now, because completely new disciplines within AI are being added on a regular basis,” says Teuwen. “But when you look at breast cancer screening, I would say the technology is at a point where we can deploy it.”
The radiologist may obtain information that is not verifiable
But Broeders and Teuwen insist on the need to think carefully about the (ethical) questions that come with AI. According to Broeders, it’s entirely possible that the algorithms could detect changes that cannot be seen by the human eye. “That means that as a radiologist you get information that you can’t control.”
“Do we trust such an algorithm? Is it acceptable to make a decision based solely on such an algorithm?” asks Broeders aloud. According to her, there are many questions that we, as a society, must answer. “Does an algorithm have to be perfect? Human work is often not perfect either, but with technology we often have the idea that it has to work perfectly.”
Teuwen also warns that most algorithms are not yet ready for widespread use. Most algorithms are trained with a bounded data set. “For example, you can create a screening algorithm on women from the Netherlands. But if we want to apply the model to the United States, we need to carefully consider whether it works as well there.”
This is also the case for the study of eBioMedicine. This algorithm performed very well in the dataset in which it was developed. Broeders: “They’ve also tested it in a new dataset. It’s already doing worse there, as it usually does. They’re also making it clear that it needs further investigation.”
“Actually, you want this scientific justification”
And then there’s the proliferation of companies developing these kinds of AI algorithms. “On the one hand, it’s great that there are so many companies doing this,” Broeders says. “But on the other hand, it is also difficult. Suppose you are a radiologist who wants to apply this, it is currently quite difficult to understand which companies offer good algorithms.”
In the study eBioMedicine is well executed and therefore reliable, explains Broeders. “But those kinds of studies don’t exist for all algorithms. And you actually want that scientific rationale.”
So there are still a lot of questions to answer before AI algorithms are actually used in cancer screening. But both Broeders and Teuwen are optimistic about the developments. Teuwen: “The techniques are very revolutionary and allow us to do things that we previously thought were impossible.”
“Bacon trailblazer. Certified coffee maven. Zombie lover. Tv specialist. Freelance communicator.”