Professor Geoffrey Hinton, the acknowledged godfather of neural network theory, was convinced in 2016 that we should stop training radiologists. Their job would soon be done by artificial intelligence. Of course, the British computer scientist and cognitive psychologist had reasons for his steep thesis. It is indeed the case that AI is good at recognising certain diseases on X-ray images.
Nevertheless, two years after his spectacular foresight, I decided to do my residency in radiology. I have now been a specialist at the Charité since my residency - and one thing has not happened so far: Too little work.
But even today, students I supervise in training ask me whether it is worth becoming a radiologist - in other words, whether they will still have a secure job tomorrow. I always reassure them: in fact, more radiologists are needed rather than fewer.
What made this renowned professor make such a prophecy? First of all, it was a bit of an exaggeration that AI could already be so far advanced in just a few years that it could safely assess all images and provide reliable diagnoses. Therefore, I am not worried that this technology will one day take over my job. In order to be able to train the algorithms sufficiently at all, you first need an incredibly huge amount of data. You simply don't have ten million images of pneumonia, for example, that the AI can learn from. It is extremely difficult to obtain this amount of data - also for data protection reasons. These are structural problems for the further development of AI in the medical field.
A second point is that before doctors can actually be replaced by AI, various ethical aspects must be discussed. The question is, for example, analogous to autonomous driving: Who is liable in the event of a misdiagnosis that leads to a patient's death? The AI programmer? The manufacturing company? The hospital?
Despite the unsolved issues, I would also state that I cannot imagine the future of medicine without AI. It will not replace us radiologists in the foreseeable future, but it will help us to take over certain routine tasks so that we can focus on more meaningful activities, such as talking to patients or presenting cases at interdisciplinary conferences. This is how AI can be used as a tool of the trade. Surprisingly, there are still only few softwares that are really proved and tested. This is because on the one hand, the data volumes for training the algorithms are lacking and, on the other hand, because there is still a lot of scepticism in the hospitals. And a part of this is also due to the costs involved in AI applications.
One example: We were offered software that could detect fractures extremely well. That is really great. But it is supposed to cost as much per year as a full time doctor's job position - but it can only do one task, namely, to recognise fractures. That doesn't seem to make sense at the moment.
The current capacities, and near-future potential of AI is something I am also researching myself at the Charité. I learned programming on the side, I can understand the algorithms, and can train and develop them for our purposes. I am also given research days for this. That's fun and makes sense.
I am currently working with coma patients after cardiac arrest. Once they are brought in by ambulance, it is very important to determine quickly if they already suffered hypoxic brain damage. That is extremely relevant for a prognosis. So I showed the AI pictures of patients who survived and those who died. It now looks for features in the images to make a prognosis along with other parameters. At the moment, we are using retrospective data to determine if the AI would have been right or wrong in real-world case. There are obvious cases where the doctor and the AI are 90 per cent right. And for the trickier cases, we are now looking for image features that the human eye does not detect, but the AI does.
My forecast is of course based on my own experience until now: The flood of medical images is increasing. We radiologists need the help of AI so that we can concentrate on the more complex tasks. We doctors will not be replaced by AI. It makes our work more exciting because we change our routines. But it goes beyond this: "old-school" radiologists who don't keep themselves technologically fit will soon be left behind.