10 Questions With….. Bram van Ginneken
Bram is a Professor of Medical Image Analysis at Radboud University Medical Center and chairs the Diagnostic Image Analysis Group. He is co-founder of Thirona, a high-tech company focusing on the development of automated medical image analysis. Bram obtained his PhD on Computer-Aided Diagnosis in Chest Radiography and went on to develop software for tuberculosis and lung screening which is now used in many countries across the world.
How would you describe what you do for a living and what impact does it have?
What I typically say to people is that my team develops computer software that analyses medical images to do the same kind of thing that doctors do. At the moment, doctors are spending a large part of their job just looking at images and saying what’s wrong with the patients. They often have to search for tiny lesions like a needle in the haystack and since we don’t have enough doctors and healthcare is extremely expensive, developing this software can have a huge impact. Automating the scanning of images we can reduce costs and also make the job more enjoyable for doctors.
How did you come to what you do?
While I was studying Physics at university I saw an advertisement for a PhD position which was about writing a computer program to take a chest x-ray and predict if someone has tuberculosis. I was not planning to do a PhD but this was a very practical project and really appealed to me. I ended up becoming a professor and leading a big research group. Developing computer algorithms that do something really intelligent doesn’t have to be in healthcare but I ended up in that area and it led me to producing tuberculosis and also lung cancer screening software which is now in active use everyday all over the world.
Why do you think there is hesitancy when it comes to AI specifically in healthcare?
In my experience, doctors are open to using these kinds of tools and if they work well they are very happy to use it. If for example a license runs out and the software is no longer they will immediately call and insist it must work again. But if the software doesn’t work so well they are very quick to say it is unacceptable. But I think this is inevitable, when a new product comes on the market the very first version is usually not fantastic. The earlier doctors that have tried it are likely to say it is useless but then in time a much better product quickly becomes available and it is then taken for granted.
Machine learning was the big buzz 10 years ago but what’s your vision for the future?
I think the next big thing is actually implementing software that works. We are definitely at the very beginning. Medicine is not a sector where things move very quickly because it is heavily regulated when talking about patients and lives. So I hope the next big thing is having successful, well working software across the healthcare system but this will take longer than we think. I also hope this will keep healthcare affordable.
I think the end goal should be to make high quality healthcare accessible to more people all over the world, which can only be achieved if we keep it affordable. I think AI can help enormously with that. If you imagine our society without computers then that is the biggest difference when you think about 50 years ago. I think 50 years from now these computers will be doing the kind of tasks we are doing now. The question is then what is left for humans but of course humans should be in charge of things and should not go out of work.
If we are able to automate the simplest of tasks will this have a huge impact on healthcare systems in terms of freeing up physicians time and reducing costs?
Yes I think most research being carried out focuses on improving the quality of healthcare and a very small part is about reducing costs. If you look at what could have a big impact in society then I believe reducing costs has a greater impact than improving healthcare. If we didn’t research how to improve healthcare that would be stupid but there is so much emphasis and prestige around it and much less research looks at reducing costs.
In some cases improving healthcare leads to reduced costs for example if you automate what doctors do today then you save costs because computers can do things much faster and have a lower salary but I think a lot of the improvements in healthcare actually increase costs tremendously. For example super expensive cancer drugs improve healthcare a little, for a small percentage of the population, but at a huge cost. So I personally think improving healthcare ultimately will not always reduce costs but increase them. But if you are willing to save lives you will be willing to pay a high price for that.
Do you think that within the healthcare sector the implementation of AI should or could be faster?
Well it really depends, some tasks are simpler and can be implemented faster but others are more complicated and therefore take longer. I think right now we should aim to implement the simple things and then hopefully society will accept that we need to develop these algorithms and there will be funding to sort out the more difficult cases. A lot can be done but it is still a lot of work and time to develop these systems and it’s a big investment one has to make. It is not that implementation is technically difficult, like some radiologists may think because their IT system has said ‘our PACS system is not compatible’. Implementation is easy for example with StrokeViewer which has been cleverly designed to be easily implemented around the current workflow and PACS system. So technically it can often be done, but there must be an incentive, demand and a funding scheme so that it is worthwhile for companies to develop this software for radiology departments to buy.
Would you be able to comment on general AI and narrow AI?
So at the moment we only have narrow AI. They are often seen as separate things but in reality there is a continuum between narrow and general AI. Eventually we will get collections of narrow AI solutions which will make up general AI, for example finding all different relevant diseases on a chest x-ray. So for us, as software developers, it will become easier and easier. Even now I can see that what I am able to do with a student in 3 months used to take us 3 years. We used to train on data sets of 100 images and are now training on data sets of 100,000 images. Facebook is training networks with 2 billion images. Things are definitely moving and developing, it is very exciting.
There has been more and more focus on prevention and early detection rather than cure, if we continue to work with narrow AI will we be able to detect multiple diseases someone may face at one time?
Well preventing someone from getting a disease is the most important but that is more to do with lifestyle and is outside the realm of what I’m doing with image analysis. My work is early detection which I think is a completely different thing. But yes, AI products developing on the market today are for the most common abnormalities because there’s a market for it. I hope in the future we will have implemented AI for the more common diseases and can develop software for more rare diseases. In some areas of early detection we will have better scans although CT scans is a fully developed technology. It has not improved for the last 15 years, they are so good they measure almost every photon of radiation you give to the patient. To get a better scan you have to give the patient more radiation which is harmful so we won’t do that. In the future, yes I think things will be totally transformed by AI. We will be able to see diseases earlier if we want to, but then we have to decide what we want to do with that. For patients, people get very scared if they hear they have early stage cancer but we all have cancer, we just don’t detect it until is something that could be serious. In the future if we get better detection mechanisms we have to deal with that and I think it’s going to be very difficult for people to handle. If there are more possibilities of detecting disease early, do you really want to know that? Is it always necessary to know that?
Can you think of which living person you admire most and why?
Very difficult question, one of the big nice things working in science is you see and interact with very clever people. In many ways more clever than I am which is very inspiring. I am using deep learning to analyse medical images. But I didn’t invent it, those that did are so clever and I think it is so fantastic what they have developed. There are a lot of people I admire in the field.
Finally, are you excited about the future of AI within medical imaging?
What I find exciting and also scary is the development of the last couple of years. We now have these AI systems that can actually create images. So for example, a simple application is a very low dose image and you improve the quality of the image. There are also systems where you take an X-ray and predict a 3D scan and it actually looks very realistic. We know this from deep fake images, you may have come across this online when someone is made to look like they are talking and it’s actually completely computer generated and looks really realistic.
I think this has great applications in medical imaging analysis. We will be able to generate complete 3D scans from very little data but it’s also very scary. We won’t know whether a computer has made up certain structures in an image or not and whether we can actually trust the data that the computer generates. I think at the moment that is one of the most exciting areas in deep learning and image analysis.
Nico.lab improves clinical practice with innovative and trusted AI. Our product StrokeViewer, is an AI-powered clinical decision support system, offering a complete assessment of relevant imaging biomarkers within 3 minutes.