Artificial intelligence is the theme at TU Delft this year. Delta explains six pieces of AI research. Part 6: using AI to read medical scans.
AI-generated image of a radiologist and a robotic assistant. (Photo: Adobe Firefly)
A group of spots and smudges in grey tints. MRI scans are not much more than this for laypeople. Just try to find a tumour, infection or haemorrhage on it. Only when the radiologist explains the image does the patient start understanding the picture.
Over the last few years, artificial intelligence has helped radiologists in hospitals all over the world interpret the images. It analyses scans faster and, in many cases, better. But this could be even better says Qian Tao. In CHEME, the AI Lab at the Faculty of Applied Sciences, the assistant professor is searching for an answer to the question how can the AI that reads the medical scans be made more trustworthy.
AI still makes mistakes, as she herself saw at her previous place of work at the Leiden University Medical Centre. “What makes it difficult is that we sometimes have no idea where the errors are made. And if you do not understand this, it is hard to trust AI at all.”
More reliable
At TU Delft, Tao is looking into how these errors are made. Her objective is to turn the artificial intelligence radiology assistant into a trustworthy partner. It is a major and popular area of research around the world.
She is trying to make AI processes more transparent. “We do not know exactly how AI sifts through the huge pile of data to arrive at a conclusion. But we want to understand this as far as we can so that we can fix any errors.” It is a huge challenge.
Tao does not think that transparency means that you have to understand everything. Another way to do it is having the AI show an uncertainty estimation with each outcome, which gains more insight in the process. “There is a big difference between 20% and 90% certainty.”
Health benefits
Apart from saving time, AI has some health benefits, says Tao. The hope is that at a certain point, AI is sensitive enough that less contrast medium needs to be given to the patient for a scan. This is important for people whose kidneys do not work well.
The artificial intelligence that reads the scans is trained using historic data. The AI assistant radiologist is thus highly familiar with common illnesses. But if a condition is rare or poorly documented, it is hard for the AI to categorise the symptoms that it sees in the medical images.
Highly specialised tasks
AI is definitely not perfect. This, emphasises Tao, is why the radiologist always has the final word. “We will not stop training radiologists. The AI is an assistant and does not take over the work. Up to now, it can only run highly specialised tasks.”
In contrast to artificial intelligence, a radiologist is a human being made of flesh and blood that can look beyond a single scan. And Tao says not to underestimate the element of human contact. “Specialists always talk to their patients before they make a diagnosis and come up with a treatment plan.”
What emerges from that, can never be quantified completely. Tao: “AI does not take the patient’s feelings and wishes into account. Therefore I do not believe that artificial intelligence will ever replace doctors.”
- Also read the other parts in this series:
- Robots that navigate like humans
- How AI simulations help to make wind turbine blades lighter
- Is it a bird? Is it a drone?
- D@S lab develops AI to benefit the unseen workers who make AI work
- A psychiatrist in your pocket: is that reassuring or worrying?
Comments are closed.